The pillar page answers seven questions briefly. This page answers sixty, in depth, grouped by theme. The goal is not to sell a viewpoint. It is to give a straight answer to every question we actually get asked in sales calls, audits and board meetings, including the ones where the honest answer is "nobody knows" or "the vendor selling that doesn't document how it works."
Every factual claim below traces back to a named source. Where something is Artlogic's practice or opinion rather than a documented fact, we say so. Where the field has not settled an argument, we describe both sides rather than pretending consensus exists.
Jump to a section
- GEO fundamentals
- How AI search actually works
- Getting started
- Content and structure
- Technical and structured data
- Crawlers and access control
- Measurement and reporting
- Risk, ethics and limits
GEO fundamentals
What is GEO, exactly?
Generative Engine Optimization is the practice of making a company legible, verifiable and consistently described across the web, so AI answer engines — Google's AI Overviews and AI Mode, ChatGPT, Gemini, Claude, Perplexity and Microsoft Copilot — can find it, trust it, and reuse it when answering a buyer's question. It rests on four things being true at once: your company is a resolvable entity a machine can distinguish from similarly named ones; independent sources corroborate your claims; your content is structured so an answer can be lifted intact; and your infrastructure actually permits retrieval, meaning the right crawlers are allowed in and pages render as readable text.
It is not one technique. It is closer to an operating discipline that spans entity data, content structure, and access control, and it compounds slowly rather than producing an overnight ranking change. See the GEO stack for how those layers depend on each other.
Where does the term "GEO" actually come from?
From a peer-reviewed paper, not a vendor. GEO: Generative Engine Optimization (arXiv:2311.09735) was presented at KDD 2024 by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. The abstract states the authors "demonstrate that GEO can boost visibility by up to 40% in generative engine responses" and notes "the efficacy of these strategies varies across domains, underscoring the need for domain-specific optimization methods." The paper also introduced GEO-bench, a benchmark of diverse queries paired with web sources, used to measure which content changes actually shifted a source's visibility in generated answers.
That study predates AI Mode and the current generation of models, so its numbers should be read as directional evidence about how synthesis systems select material, not as a live 2026 ranking-factor list. We use them that way throughout this hub.
Is GEO the same thing as AEO or LLMO?
There is no industry consensus, and it is worth knowing that before you sit across from a vendor who insists otherwise. GEO, AEO (Answer Engine Optimization), LLMO, AI SEO, GAIO and "AI search optimization" are all in circulation for roughly the same set of activities. Profound, a vendor in this space, publicly argues "GEO" is a poor term because it collides with geo-targeting, and pushes "AEO" instead. Other vendors use the terms interchangeably. Wikipedia lists them as related without differentiating.
Our position: the label matters far less than the underlying discipline. We use "GEO" because it is the term from the original academic paper, not because it has won any naming argument. If an agency's entire pitch depends on convincing you their term is the correct one, that is a sign to look past the vocabulary and ask what they actually propose to do.
Does GEO replace SEO?
No, and be skeptical of anyone who tells you otherwise. Google still delivers roughly 500 times more referral traffic than ChatGPT from search alone, and roughly 1,300 times more including Discover, according to Reuters Institute's 2026 analysis. Traditional search is still the overwhelmingly larger channel. GEO shares SEO's eligibility layer — a page has to be indexed and snippet-eligible before it can appear in any AI feature — and adds entity disambiguation, cross-web consensus, and passage-level extractability on top.
The honest framing is overlay, not migration. You keep investing in the search infrastructure that produces the clicks you already get, and add the work that produces mentions inside generated answers. See the full SEO versus GEO comparison for where the two diverge.
Is GEO a real discipline, or something vendors invented to sell services?
Both things are true, which is an uncomfortable but accurate answer. The underlying phenomenon is real and peer-reviewed: retrieval-augmented generation exists, query fan-out is documented by Google itself, and the KDD 2024 paper measured, with a benchmark, that specific content changes shift a source's visibility in generated answers. That is not marketing. It is published research with a methodology you can read.
What is also true is that a large amount of current "GEO" commentary is repackaged SEO advice, speculation dressed as fact, or outright wrong — recommending FAQ schema for Google rich results a year after Google discontinued them, for instance. The discipline is real. A meaningful share of what gets sold under its name is not evidence-based. Judge any specific claim on its source, not on whether the word "GEO" appears near it.
Who actually needs a GEO programme?
Companies whose buyers research a considered purchase conversationally before contacting anyone. That is most B2B software, professional services (legal, accounting, consulting), healthcare and clinical services, financial services, and higher-value real estate and relocation decisions — categories where a buyer plausibly asks an AI tool "who are the best X for Y" and forms a shortlist from the answer. Reuters Institute's 2026 data shows AI Overviews already appear in roughly 10% of US search results and ChatGPT has roughly 800 million weekly active users, so this audience is not hypothetical.
It matters less for businesses selling on price or immediacy in a fixed local market, where the buying decision is unlikely to route through a synthesized answer at all. If your buyers Google "plumber near me" and call the first result, GEO is a lower priority than making sure your Business Profile and phone number are correct.
Who can reasonably skip GEO for now?
Two groups. First, businesses whose demand is almost entirely local, urgent, or price-driven, where buyers act on the first usable result rather than building a shortlist from a synthesized answer. Second, and more importantly, any organization that has not yet fixed its basics: if your site is not indexed, key content sits behind JavaScript that never renders as text, or your robots.txt already blocks ordinary crawling, you are what the GEO maturity model calls Level 0, and GEO-specific work will not move anything until that is fixed.
Skipping GEO should be a deliberate choice based on how your buyers actually search, not a default because nobody assigned the work. Revisit the decision periodically — Reuters Institute's 2026 data shows AI Overview prevalence rising, and a market that does not need this today may need it in eighteen months.
What does it cost a company to do GEO badly?
Three distinct costs, and they are not hypothetical. Money: paying for llms.txt implementation or FAQ schema as visibility tactics that no major vendor documents as consumed or rewarded — llms.txt has no vendor adoption, and FAQ rich results stopped appearing in Google Search on 7 May 2026. Visibility: blanket-blocking AI crawlers to protect against training use, without realizing that OAI-SearchBot and PerplexityBot control whether you appear in ChatGPT and Perplexity search, not training. And in the worst case, existing traffic: Otterly.ai's experiment publishing roughly 1,000 AI-generated posts each on two fresh domains ended with both sites algorithmically deindexed by Google, one falling from 1,629 to 15 daily impressions. That is a two-site sample, not a rule, but it is a real cautionary data point, not a hypothetical one.
How AI search actually works
What is "query fan-out"?
It is the technique Google documents for AI Overviews and AI Mode: "issuing multiple related searches across subtopics and data sources" rather than searching for the literal words a person typed. AI Mode goes further, according to Google's own description — it "makes a plan, conduct[s] searches to find information and adjust[s] the plan based on what it finds," running several searches concurrently across subtopics.
The practical consequence is that optimizing a page for the exact phrase a buyer might type is the wrong target. A prompt like "who should we hire to fix our AI search visibility" gets decomposed into several sub-queries the engine invented on its own — something closer to "GEO agency Canada," "AI visibility consultant," "GEO vs SEO," each searched independently. You need to be a credible answer to those sub-questions, on pages that actually exist, not to the buyer's literal sentence. See query fan-out made concrete for a worked example.
What's the difference between parametric memory and retrieval?
Parametric memory is what a model absorbed during training and can recall without searching anything. Retrieval is the model running a live search and reading what comes back before answering. They are influenced by completely different work and on completely different timelines.
You affect parametric memory only over long horizons, by having existed prominently and consistently across the material a model was trained on — it is why long-established brands get named in generic prompts with no search step at all, and why a two-year-old company almost never does. You can affect retrieval this quarter: fixing robots.txt access, publishing an extractable answer, or correcting a third-party profile can change what a live search surfaces within weeks.
A useful diagnostic: if a model names you in a prompt with no browsing indicator, that is parametric memory working. If it shows sources or a "searching the web" step, that is retrieval, and that is the lever you can actually pull.
flowchart TD
A["Buyer asks a question"] --> B{"Does the model decide it needs current information?"}
B -->|"No"| C["Answer drawn from parametric memory"]
B -->|"Yes"| D["Query fan-out into sub-questions"]
D --> E["Retrieve and rank candidate pages"]
E --> F["Synthesize an answer"]
F --> G["Attach citations"]
C --> H["Response reaches the buyer"]
G --> H
Two entirely different routes produce the same response. Route one is shaped by what existed in training data years earlier. Route two is shaped by what is retrievable today.
Why do I get a different answer from the same AI tool on different days?
Because generated answers are genuinely non-deterministic, not because something is broken. The same prompt can produce different responses across sessions, across users, and across silent model version updates, and none of the major vendors publish a changelog granular enough to explain a specific shift. Sampling adds another layer of variance: Evertune's measurement approach runs each prompt 100 times across eleven models specifically because a single check tells you almost nothing about the underlying rate.
This is the single biggest reason ad-hoc "I typed our name into ChatGPT and we weren't there" checks are unreliable evidence of anything. One query, one session, is an anecdote. A defensible read on your visibility requires a fixed prompt set, sampled repeatedly, tracked over time, per engine — which is exactly the discipline behind share of answer measurement.
How does each AI engine actually retrieve information?
Differently enough that treating them as one channel produces bad diagnoses. Google's AI Overviews and AI Mode use query fan-out across the existing Search index and Knowledge Graph. ChatGPT search "rewrites your query into one or more targeted queries" and sends them to search partners that OpenAI's help documentation names as including Bing and Shopify, without stating either is exclusive. Gemini's grounding documentation describes the model deciding a search is needed, generating and running queries, then returning inline url_citation annotations with character offsets. Claude, per Anthropic's documentation, decides autonomously whether to search, can search multiple times in one turn, and always attaches citations as web_search_result_location objects. Perplexity performs real-time retrieval with citations on every answer, though its ranking backend is not publicly documented. Copilot is "generally understood" to inherit Bing's index — Microsoft has not published a current architecture statement we could verify, so we say "generally understood" rather than asserting it as documented fact.
Do AI models really "read" my website, or just a cached snapshot?
Both happen, and they are different mechanisms with different guarantees. Ordinary crawling and indexing produces a cached, sometimes stale, representation used for training or for building a search index — this is what Googlebot, GPTBot or ClaudeBot are doing. Retrieval-time fetching happens live, at the moment a model decides to search, and reflects whatever the page contains right now.
There is a third path worth knowing about, because it is the one path where your page is guaranteed to be read in full: a person pastes your URL or uploads your PDF and asks the model to look at it directly. Agents like ChatGPT-User, Claude-User and Perplexity-User exist for exactly this, and it makes on-page clarity a direct sales asset in that moment, not just a longer-term visibility one. Beyond that guaranteed case, whether and how often a given page gets fetched live is not something any vendor documents in detail.
What happens when different sources say different things about my company?
Generative systems function as consensus estimators: when independent sources largely agree on a claim, that claim becomes retrievable as fact. When sources conflict, the model does one of three things — hedges the answer, favors whichever source it judges most authoritative, or omits the entity entirely to avoid asserting something it cannot support. None of the major vendors publish the exact logic that decides between those three outcomes, so this is inference from behavior, not documented mechanics.
The practical implication is that inconsistency is not a neutral state. A company whose website, LinkedIn, Crunchbase and press mentions describe it four different ways is not "covered from four angles" — it is a source of conflicting signal that a consensus machine may simply resolve by leaving the company out. Fixing this is unglamorous, cheap, and one of the highest-yield hours in most engagements: pick one canonical description and repeat it everywhere.
Does freshness matter, and how fresh does content need to be?
Retrieval-time answers pull from current search indexes, so a page that changed an hour ago can, in principle, be reflected in a live search-backed answer the next time a fan-out query touches it — that is the whole point of Google's fan-out approach and of IndexNow, the protocol Bing and several other engines use to receive instant change notifications. Parametric memory is the opposite: it is frozen at training time and updates only when a model is retrained, on a schedule no vendor publishes.
No major vendor documents a specific "freshness" ranking factor for AI features beyond the freshness signals that already apply to ordinary Search indexing. Be wary of anyone quoting a precise re-crawl or re-training cadence — it is not published, and it likely varies by engine and by page. The honest operational answer is: fix something, wait weeks not days, and re-measure with a fixed prompt set rather than guessing at a timeline.
How does an engine decide which sources to cite in an answer?
Partially documented, and partially not, which is worth saying plainly rather than filling the gap with speculation. Gemini and Claude document the citation mechanics precisely — url_citation objects with character offsets, and web_search_result_location objects with literal cited text — but neither vendor publishes the ranking logic that decides which retrieved candidates make the cut. Perplexity attaches citations to every answer, but its backend ranking is explicitly not publicly documented. Google states AI Overviews and AI Mode may run query fan-out and draw on the Knowledge Graph and real-world data, without specifying a citation-selection algorithm.
What is observable rather than documented: Profound's analysis of 11.84 billion citations found roughly 57% point to brand-owned domains, and Semrush's study found AI Overviews carry around 11 supporting links on average with only 20–26% overlap with top-ten organic results. Those are patterns in outcomes, not confirmed mechanisms — treat any confident description of "the algorithm" as speculation dressed as fact.
Why can I rank #1 on Google and still never get cited by an AI engine?
Because ranking and citation are answering different questions. Semrush's study of 200,000 US keywords found that more than half of desktop AI Overviews did not link the #1 organic result, and the overlap between AI Overview links and the organic top ten was only 20–26%. Ranking first tells you a page is relevant and authoritative for a query. It does not tell you whether the page is easy to lift a self-contained answer from, corroborated by independent sources, or actually reachable by the retrieval agents involved.
The gap is usually one of three things: the page answers the query well but buries the answer past the first few paragraphs, so nothing is cleanly extractable; the entity behind the page is ambiguous or inconsistently described elsewhere, so the system has less confidence citing it as a fact source; or a retrieval agent specific to that engine — OAI-SearchBot, PerplexityBot, Claude-SearchBot — is blocked while Googlebot is not. Rank is necessary but not sufficient.
Getting started
What's the single first action I should actually take?
Read your own robots.txt file and identify every AI-related user-agent listed in it, then work out which decision each line actually makes. In a meaningful share of the audits we run, this one file explains most of a company's invisibility in AI answers, usually because someone blocked crawlers broadly to protect against AI training use without realizing that some of those same tokens control retrieval and search visibility, not training.
The check takes minutes. Load yoursite.com/robots.txt, list every disallowed user-agent, and separate them into two buckets: training crawlers (GPTBot, ClaudeBot, Google-Extended) and search or retrieval crawlers (OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot, bingbot). If a search or retrieval crawler is disallowed, that is very likely costing you visibility on that specific engine, and it is usually the cheapest fix available to you.
What should I check before spending any money on GEO?
Four things, in order, because spending on entity work or content while access is broken wastes the budget. First, is the site indexed and snippet-eligible in ordinary Google Search — Google states this is the same eligibility gate AI Overviews and AI Mode use, with no additional requirement. Second, does robots.txt allow the retrieval agents relevant to the engines your buyers actually use. Third, does the important content on your key pages render as visible text, not JavaScript-only or locked inside images and PDFs. Fourth, is your company described the same way across your own site, LinkedIn, Crunchbase, and any directories that already carry your information.
If any of those four are broken, fix them first. They are inexpensive relative to content or authority-building work, and nothing built on top of a broken foundation performs — this is the GEO maturity model's Level 0 and Level 1 distinction in practice.
How do I audit my robots.txt for AI crawlers specifically?
Pull the live file, then go line by line against a reference table of documented agents — the one on the GEO pillar page is a reasonable starting point. For each disallowed user-agent, ask what it actually does: GPTBot collects training data and has no effect on ChatGPT search visibility if blocked; OAI-SearchBot surfaces you in ChatGPT search and removing it removes you from that surface; PerplexityBot does the equivalent for Perplexity; Google-Extended opts out of AI training and grounding in some Google systems without touching Search indexing; Googlebot disallowed removes you from Search entirely, including AI Overviews and AI Mode.
Then verify the crawlers you do allow are actually reaching you — check server logs or a CDN's bot-traffic report for hits from the documented IP ranges (OpenAI publishes theirs at openai.com/gptbot.json and openai.com/searchbot.json; Anthropic at claude.com/crawling/bots.json) to confirm nothing at the network layer is silently returning errors to legitimate requests.
Do I need to hire an agency for this?
Not for all of it. The access and entity fixes — correcting robots.txt, unifying your company description across properties, adding basic Organization structured data — are within reach of an in-house marketing or engineering team with a half-day and a checklist. Where outside help earns its cost is usually the parts that require ongoing measurement discipline and cross-web relationship building: baselining share of answer across engines with a properly sampled prompt set, and earning independent citations at a defensible pace rather than in a way that reads as manufactured.
Be skeptical of any agency, including ours, that frames the entire discipline as something only they can execute. A fair test of any pitch: ask what specifically they will measure, on which engines, against which named competitors, and how they will tell improvement from ordinary variance. If a structured audit is the right next step for your situation, that conversation should start with a baseline, not a contract.
What does a minimum viable GEO programme actually look like?
Four things, done in sequence, without a large budget. One, a robots.txt fix separating training decisions from retrieval decisions. Two, a single canonical one-sentence description of the company, updated to match across the website, LinkedIn, and any directories that already carry your listing. Three, one answer-first page rewritten or created for a real, specific buyer question — the answer in the first sixty words, elaboration after. Four, a fixed set of ten to twenty prompts your actual buyers would plausibly ask, checked across two or three engines at the start and again a month later.
That is not a complete programme. It will not produce a defensible cross-engine measurement or build the cross-web consensus that compounds over years. But it costs almost nothing, it fixes the two most common unforced errors, and it gives you a real baseline instead of a guess.
We just launched. What does GEO look like for a brand-new domain?
Slower and more retrieval-dependent than for an established brand. A new domain has no presence in any model's parametric memory, so you cannot be named by a model with no search step — every appearance you get will come through live retrieval, which makes the access and extractability basics disproportionately important early. There is also no accumulated corroboration yet: no third-party mentions, no consistent multi-source description, nothing for a consensus-estimating system to confirm.
Practically, that means front-load the identity work — a clean Organization entity, a sameAs set pointing to every verified profile, one description used everywhere from day one — before content volume. A new domain publishing at scale without that foundation is also the exact profile of the two sites in Otterly.ai's deindexing experiment, so build slowly and verifiably rather than fast and thin. Expect this to be a multi-quarter effort; nothing here shortens the parametric-memory timeline.
What has to already be true before GEO work can even start?
The ordinary Search eligibility layer. Google states plainly that the gate for AI Overviews and AI Mode is that a page be "indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements" — no additional technical requirement exists on top of that. In practice this means: the site is crawlable, not blocked by robots.txt or a misconfigured CDN; important content is server-rendered or otherwise visible as text, not dependent on client-side JavaScript that a crawler won't execute; there are no noindex tags on pages you want found; and internal links make your content actually reachable rather than orphaned.
If any of that is broken, no amount of entity work, schema, or content restructuring will produce visibility, because the page is not in the pool the retrieval systems draw from in the first place. This is the diagnostic starting point for the flowchart below.
Content and structure
Do I need to write content differently for AI than for humans?
Not fundamentally. You need to write content that answers a real question clearly, in text, near the top of the page, supported by specifics — which happens to be what generative retrieval rewards and also what a human reader wanted all along. The KDD 2024 GEO paper's directional findings support this: keyword stuffing, the archetypal "written for a machine" tactic, measurably reduced visibility in generated answers, while quotations from experts and statistics, both markers of genuinely useful reference material, produced the largest measured gains.
The one real adjustment is structural rather than stylistic: put the direct answer to the implied question in the first few sentences rather than building up to it, because an extraction system is more likely to lift a self-contained early passage than to read fourteen paragraphs deep. That is good writing practice regardless of any AI system reading it.
What does "answer-first" structure actually mean?
It means the direct answer to the question implied by a heading or a page appears in the first sentence or two beneath it, with supporting detail, caveats and nuance following afterward rather than preceding it. A paragraph that opens with three sentences of context and lands on the actual answer in sentence four is much harder for an extraction system to lift cleanly, and it is also harder for a human skimming the page to use.
Practically: write the heading as the question a buyer would actually ask, answer it plainly in the first sixty to eighty words, then elaborate. This is the pattern the answers on this page are built to demonstrate, not just describe. It maps directly to what the GEO paper found about extractability — the tactics that helped were the ones that made a passage quotable in isolation, and an answer buried under throat-clearing is not quotable in isolation.
Should content be long or short to get cited?
Length itself was not one of the nine tactics the KDD 2024 GEO paper measured, so there is no controlled evidence that length alone changes citation likelihood in either direction. What the paper did measure — quotations, statistics, fluency, source citations — are things that can happen inside a short passage or a long one; they are about density and quotability, not word count.
Separately, Semrush's study found that roughly 82% of AI-Overview-triggering keywords had fewer than 1,000 monthly searches and about 80% were informational in intent, which suggests the content that wins these queries tends to be specific and explanatory rather than broad commercial copy — but specific does not mean short. A thorough, well-sourced page that answers a narrow question completely can outperform a shallow short page and a padded long page equally. Optimize for a complete, quotable answer to a specific question, and let length follow from that rather than targeting a number.
Is it safe to publish AI-generated content at scale?
Not demonstrably, and there is a real cautionary data point rather than just a general worry. Otterly.ai ran a controlled experiment publishing roughly 1,000 AI-generated posts each on two fresh domains. Both were algorithmically deindexed by Google, with no manual action notice and no appeal path; one site's daily impressions collapsed from 1,629 to 15. That is a sample of two sites in one niche, and it should not be treated as proof that all AI-assisted content triggers deindexing — plenty of sites use AI drafting responsibly, with human editing and genuine expertise layered in. But it directly falsifies the confident claim that scaled, unedited AI publishing carries no risk.
The reasonable middle position: use AI tools in the drafting process if that helps you move faster, but keep a human accountable for accuracy, add genuine expertise or data the model could not have generated, and do not treat publishing volume as a strategy in itself.
Should I add quotations and statistics to my pages?
The evidence for this is more specific and more dated than most people repeating it realize. The KDD 2024 GEO paper measured nine content modifications against its own benchmark: adding quotations from expert or relevant sources produced the largest single measured gain, roughly 40%; adding statistics and data points produced roughly 30%; citing sources produced roughly 27–28%. Those numbers should be read as directional findings from a study that predates AI Mode and the current model generation, not as a live ranking-factor table you can apply mechanically in 2026.
The underlying reason still holds regardless of the exact percentage: a sentence with a real number in it or a claim attributed to a named expert is easier to extract and quote in isolation than a hedged, generic paragraph. That is a durable property of what makes text useful reference material, which is why it is worth doing even without treating the 2024 figures as current.
Do comparison pages actually help visibility?
There is supporting evidence that comparison and "best X" content matters, though it is more about where that content lives than a guarantee for your own site. Semrush's study found this content type dominates the queries that trigger AI Overviews, and the pillar page's authority section notes that appearing in third-party comparison and listicle content is one of the stronger observed authority signals — precisely because it is independent corroboration rather than a company describing itself.
A comparison page you publish yourself can still help, mainly by giving a clear, fair answer to a real fan-out sub-query buyers generate, but it does not carry the same corroborating weight as the same comparison appearing on a trade publication or an independent review site. Neither outcome — appearing in third-party comparisons or ranking well for your own — is documented as guaranteed by any vendor. Treat comparison content as one input into consensus, not a citation-earning mechanism on its own.
Should I name competitors directly on my own site?
No vendor documentation states this helps or hurts citation likelihood specifically, so treat any confident claim either way as opinion. What is documented is that comparison-shaped queries are common in the sub-questions engines generate through fan-out, and that third-party comparison content is heavily represented in AI answers. A fair, specific, on-site comparison can be a genuine answer to one of those sub-queries.
The practical risk is not algorithmic, it is credibility: a comparison page that reads as self-serving or unfair is a weaker corroborating source than an independent one, and it can undercut the authoritative tone that the GEO paper found to have a modest positive effect. If you name competitors, do it accurately and let the differences speak for themselves rather than editorializing. If you are not confident you can be fair, an independent comparison you did not write is worth more to your visibility than one you did.
How often should existing content be refreshed?
No vendor documents a specific refresh cadence tied to AI visibility, so any number you hear quoted as a rule is not sourced from anywhere official. What is documented and worth building a cadence around instead is link health: Otterly.ai found 19.3% of over 20 million cited URLs across seven AI engines were dead in a one-month snapshot, with ChatGPT the highest at 25.1% and Google AI Overviews the lowest at 12.6%. A citation you earned and then broke — through a site migration, a URL change, or a deleted page — is a citation you no longer have, regardless of how good the original content was.
A reasonable practice, not a documented requirement: review high-traffic and high-citation pages quarterly for factual accuracy and dead internal or outbound links, and treat any domain migration as a redirect-mapping project, not an afterthought.
Technical and structured data
Does adding schema markup help me get cited by AI?
Not in the way it is often sold. Google's own AI-features documentation states directly: "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add." That is unambiguous, and it is the single most important framing fact in this entire hub — there is no secret tag.
Where structured data still earns its cost is disambiguation, not ranking: Organization markup with a complete sameAs list helps a system resolve which entity you are when your name is ambiguous, and Google's guidance elsewhere stresses that structured data should match your visible text, which supports fact confirmation rather than eligibility. Add schema because it makes your entity clearer, not because a vendor promised it produces citations. It does the former reliably and has never been documented to do the latter.
Is FAQ schema still worth implementing in 2026?
Not for the reason most people implement it. Google's documentation states plainly: "As of May 7, 2026, FAQ rich results are no longer appearing in Google Search. We will be dropping the FAQ search appearance, rich result report, and support in the Rich results test in June 2026. To allow time for adjusting your API calls, support for the FAQ rich result in the Search Console API will be removed in August 2026." HowTo rich results were removed earlier still, in September 2023. Anyone recommending FAQPage markup as a Google visibility tactic in 2026 is working from outdated information.
FAQPage markup is not worthless — it remains legitimate machine-readable semantics that non-Google consumers of structured data can parse, and shipping it costs little. Just do not expect a rich result from it, and do not pay a premium for it as a "GEO tactic." This page discusses its own schema choices in the schema note below.
Should I create an llms.txt file for my site?
You can, but be clear-eyed about what you are actually doing. llms.txt is a September 2024 proposal by Jeremy Howard of Answer.AI, published at llmstxt.org, suggesting a markdown file that gives a curated, language-model-friendly summary of a site. It has been adopted by some documentation tooling. No major AI vendor — not OpenAI, Google, Anthropic, Microsoft, or Perplexity — has published documentation stating that its crawlers or answer engines actually consume it. Google's AI-features guidance explicitly states you "don't need to create new machine readable files."
Shipping one costs roughly an hour of engineering time and does no harm. Paying a vendor for it as a dedicated visibility service is paying for an unproven hypothesis, not a documented mechanism. If your team wants to ship it as a low-cost, low-risk documentation aid, there is no reason not to. Just don't expect it to move a citation metric, and don't let anyone bill it as though it will.
Does JavaScript-rendered content hurt my AI visibility?
It can, and Google's own guidance is direct about the fix rather than the problem: keep important content in textual form, retrievable without depending on client-side rendering. If a crawler — whether Googlebot, GPTBot, OAI-SearchBot, or any other documented agent — cannot execute your JavaScript reliably, or executes it more slowly or incompletely than a browser would, the content inside it may never make it into an index or a retrieval-time fetch at all.
This is a known, mundane failure mode, not a mysterious one: the fix is server-side rendering, static generation, or at minimum ensuring core textual content is present in the initial HTML response rather than injected entirely after page load. Test it directly — fetch your key pages with a plain HTTP request and no JavaScript execution, and check whether the answer you want cited is actually present in that raw response. If it is not, no amount of content strategy fixes it until it is.
What about important information locked inside PDFs or images?
It is a genuine extractability problem, separate from and in addition to the JavaScript issue. Text inside an image has no machine-readable text layer unless OCR has been applied and exposed, and even well-structured PDFs are harder for many crawling and retrieval pipelines to parse cleanly than an HTML page — tables, multi-column layouts and embedded charts routinely lose structure in extraction. If your key facts — pricing, credentials, a named methodology, a statistic you want cited — exist only inside a PDF or an infographic, you have made them substantially harder to retrieve and quote, even if a human reader finds them easily.
The practical fix is duplication, not abandonment of PDFs and images: keep the PDF or graphic for the audiences who want it, but also publish the same key facts as plain HTML text on the page itself. That satisfies Google's stated guidance to keep important content in textual form, and it gives every retrieval agent a text-based version to work with regardless of how well it parses the original file format.
What is IndexNow and is it worth using?
IndexNow is a protocol that lets a site push instant notifications of new or changed pages to participating search engines, instead of waiting for a routine crawl. The engines currently listed as participating on indexnow.org are Microsoft Bing, Naver, Seznam.cz, Yandex and Yep. Google is notably not among them — IndexNow does not accelerate Google indexing.
It is worth adopting specifically because of the Bing connection: Bing is a named search partner for ChatGPT search, per OpenAI's help documentation, and Microsoft Copilot is generally understood to inherit Bing's index, though Microsoft has not published a current architecture statement confirming that in detail. Faster Bing indexing therefore plausibly benefits two AI surfaces at once through one integration, at low engineering cost. It is not a general AI-visibility solution and does nothing for Google, ChatGPT's non-Bing signals, Gemini, Claude, or Perplexity.
Can our own CDN or bot-protection service accidentally block AI crawlers?
Yes, and it is a more common failure than teams expect, because bot-mitigation rules are usually tuned against scraping and credential-stuffing traffic, not written with a documented list of legitimate AI crawlers in mind. Aggressive rate limiting, JavaScript-challenge pages, or overly broad user-agent blocklists at the CDN layer can return 403 responses to GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot or bingbot just as readily as to a malicious scraper, and nothing in a standard analytics dashboard will flag that distinction for you.
The check is mechanical: pull your CDN or WAF logs, filter for the documented user-agent strings and IP ranges each vendor publishes — OpenAI at openai.com/gptbot.json and openai.com/searchbot.json, Anthropic at claude.com/crawling/bots.json — and confirm those requests are returning 200s, not 403s or challenge pages. Do this after any CDN or security vendor change, since bot-rule updates are a common, silent cause of a sudden drop in AI-engine visibility.
Does Bing actually matter if most of our traffic is from Google?
More than most teams assume, because Bing's role in the AI-search ecosystem is broader than its share of direct search traffic suggests. OpenAI's help documentation names Bing among ChatGPT search's partners, without stating it is exclusive, which means Bing indexation plausibly affects ChatGPT visibility. Microsoft Copilot is generally understood to inherit Bing's index. And Bing co-created IndexNow, giving it a fast-indexing advantage most sites never activate.
Because most organizations concentrate nearly all technical SEO effort on Google, Bing Webmaster Tools is systematically under-managed — verifying ownership, submitting a sitemap, and adopting IndexNow typically takes under an hour and is comparatively uncontested territory. It will not move Google, Gemini, Claude or Perplexity visibility, but for ChatGPT and Copilot specifically, it is one of the higher-leverage, lower-effort items available. Consider a technical visibility review if you have never checked Bing indexation directly.
Crawlers and access control
What's the difference between GPTBot and OAI-SearchBot?
They are both operated by OpenAI and are frequently confused, which is exactly why this distinction matters more than almost anything else on this page. GPTBot collects data used for model training. Blocking it is a content-licensing decision with no documented effect on whether you appear in ChatGPT search. OAI-SearchBot is the agent that surfaces sites in ChatGPT search results — blocking it removes you from that surface entirely. A third agent, ChatGPT-User, handles user-triggered fetches, described by OpenAI as non-automatic rather than a background crawler.
The exact documented user-agent string for the search crawler is Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/131.0.0.0 Safari/537.36; compatible; OAI-SearchBot/1.4; +https://openai.com/searchbot. If your robots.txt disallows OAI-SearchBot alongside GPTBot under a single blanket "block OpenAI" rule, you have very likely removed yourself from ChatGPT search without meaning to.
Should I block AI crawlers that train models?
That is a legitimate content-licensing decision, and it is genuinely yours to make — there is no obligation to allow your content into a model's training data, and doing so does not affect ordinary search visibility in any documented way. Blocking GPTBot, ClaudeBot, or using the Google-Extended opt-out token are all recognized, vendor-supported mechanisms for exactly this purpose, and each is independent of your Search or AI-answer visibility.
The important discipline is keeping this decision separate from your retrieval decision. Block training agents for licensing reasons if that reflects your position on how your content should be used, but verify you have not also blocked OAI-SearchBot, PerplexityBot, or Claude-SearchBot in the same rule — those control whether you appear in ChatGPT search, Perplexity, and Claude's search results respectively, and blocking them is a visibility decision, not a licensing one, even though it is often made by the same well-intentioned person editing the same file.
What's the actual trade-off between blocking for licensing reasons and staying visible?
There is no trade-off if the two decisions are kept separate, because they use different agents — the trade-off only exists if you conflate them. You can block GPTBot (training) while allowing OAI-SearchBot (ChatGPT search visibility), and you lose nothing on the visibility side while still withholding your content from that model's training corpus. The same separation applies to Anthropic's ClaudeBot versus Claude-SearchBot, and to Google's Google-Extended versus Googlebot.
The real trade-off only appears with Perplexity, imperfectly: PerplexityBot is documented as not crawling for model training at all, so blocking it purely removes you from Perplexity's search surface with no licensing benefit attached. If your objection is specifically to training use, blocking Perplexity's crawler achieves nothing toward that goal and only costs you visibility on that engine. Know which agent does which job before writing the rule.
flowchart TD
A["A crawler user-agent shows up in your logs or you're editing robots.txt"] --> B{"What is your actual objection?"}
B -->|"I don't want my content used to train a model"| C{"Which vendor?"}
C -->|"OpenAI"| D["Disallow GPTBot only"]
C -->|"Anthropic"| E["Disallow ClaudeBot only"]
C -->|"Google"| F["Use the Google-Extended token"]
B -->|"I want to disappear from that engine's answers"| G["Disallow the matching search agent: OAI-SearchBot, PerplexityBot, or Claude-SearchBot"]
B -->|"I'm not sure, someone just said block all bots"| H["Stop. Identify the agent against a documented list before touching robots.txt"]
D --> I["ChatGPT search visibility unaffected"]
E --> J["Claude search visibility unaffected"]
F --> K["Search indexing unaffected"]
G --> L["You are now absent from that specific engine's answers"]
Training and retrieval are controlled by different user-agents. The costly mistake is treating "block AI" as one decision when it is really several.
Why does Perplexity-User reportedly ignore my robots.txt?
Because Perplexity's own documentation states that Perplexity-User generally ignores robots.txt, on the reasoning that the fetch is user-initiated rather than an automated crawl — a person directed the assistant at a specific page, similar in spirit to a browser extension fetching a page a human asked it to open.
We report Perplexity's documented position here rather than arguing about whether that policy is appropriate. Whether user-triggered fetches should be exempt from robots.txt is a live question for the industry, and this hub's job is to tell you what is documented rather than to referee it. What is actionable either way: PerplexityBot, the separate agent used for Perplexity's own indexing and citation surfacing, does respect standard robots.txt directives, so your control over whether Perplexity indexes and cites you generally still functions through that agent.
What does Google-Extended actually control?
It is Google's documented robots.txt token for opting out of having your content used for AI training and grounding in some of Google's other systems, separate from Googlebot, which controls ordinary Search crawling and indexing. Google states this explicitly, and the brief's fact base is direct on the point: Google-Extended "does not affect Search indexing."
That means you can disallow Google-Extended while still allowing Googlebot, and your ordinary Search visibility — including eligibility for AI Overviews and AI Mode, which run on the same indexing gate — remains unaffected. This is the cleanest example among all documented crawler tokens of the training-versus-retrieval split done exactly right by the vendor's own design: one token for the licensing decision, a completely separate one for the visibility decision, with no overlap between them.
How do I verify that a crawler claiming to be GPTBot is actually legitimate?
Check it against the vendor's own published verification data rather than trusting the user-agent string alone, because a user-agent header can be spoofed by anyone. OpenAI publishes the IP ranges its crawlers use at openai.com/gptbot.json for GPTBot and openai.com/searchbot.json for OAI-SearchBot. Anthropic publishes equivalent data for ClaudeBot at claude.com/crawling/bots.json. Cross-reference the requesting IP address in your server or CDN logs against those published ranges; a request claiming to be GPTBot from an IP outside OpenAI's published range is not GPTBot.
For reference, the documented GPTBot user-agent string is Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; GPTBot/1.4; +https://openai.com/gptbot, and Perplexity's documented string is Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; PerplexityBot/1.0; +https://perplexity.ai/perplexitybot). Use the string to filter logs quickly, then confirm with the IP list before treating a request as authoritative traffic.
Measurement and reporting
How do you actually measure AI visibility?
With a defined metric set, sampled repeatedly, per engine, because no single tool gives an authoritative cross-engine view. The core metrics: share of answer, the percentage of responses across a fixed prompt set that name your brand, tracked per engine because variance between engines is large; citation rate, the percentage that link you, distinct from naming you; prompt coverage, how many of your mapped buyer prompts you appear for at all; mention sentiment and framing, since being named as a leading option and being named as a cheaper alternative to a competitor are different commercial outcomes; competitive share against named competitors on the same prompts; citation health, the share of your earned citations that still resolve, given the roughly 19.3% cross-engine dead-link rate Otterly.ai observed; and, cautiously, assisted conversions.
There is no Search Console equivalent for generative answers, which is precisely why method matters more here than in ordinary SEO reporting. See Phase 05 of the methodology for how this is operationalized.
What is "share of answer" and why does it matter more than rank?
Share of answer is the percentage of responses, across a fixed set of prompts, that name your brand — measured per engine, sampled repeatedly, because the underlying rate varies with model version and session. It matters more than rank in this context because rank has no direct equivalent in a generated answer: there is no position ten, no page two. A brand is either named in a given response or it is not, and the interesting number is how consistently that happens across a realistic prompt set, not where you sit in a list that does not exist.
It is also the metric that makes cross-engine comparison meaningful, since "we rank well" collapses six genuinely different systems into one number, while "we appear in 40% of ChatGPT responses to this prompt set and 15% of Perplexity's" tells you exactly where the gap actually is and which engine's mechanics to investigate next.
Why is checking a prompt once basically useless?
Because generated answers are non-deterministic by design, and a single check cannot distinguish a real absence from ordinary variance. The same prompt, run twice in the same session, can return meaningfully different answers, and Evertune's approach — sampling each prompt 100 times across eleven models — exists specifically because that variance is large enough to distort a single-run result into a false conclusion in either direction.
If you type your company's name into ChatGPT once, get no mention, and conclude you have a visibility problem, you may be right, or you may have sampled an unlucky run. The only way to tell the difference is repetition: a fixed prompt set, checked multiple times per engine, tracked over a baseline period before you change anything. Any report — from us or anyone else — that presents a single screenshot of a single response as evidence of a trend should be treated as an anecdote, not a measurement.
What are the limits of the AI-visibility tools on the market?
The most important limit is one almost no vendor states clearly: there is no independent third-party accuracy audit of any of these tools that we have been able to verify. Ahrefs Brand Radar, Otterly.ai, Profound, Peec AI, Evertune, Scrunch, Rankscale and Semrush AI Visibility each report on their own coverage and their own sampling methodology, and each covers a different, non-identical set of engines — Peec AI covers only ChatGPT, Perplexity and Gemini, for instance, while Ahrefs Brand Radar covers seven platforms plus social sourcing.
Pricing also varies widely and changes: Ahrefs Brand Radar runs $398–699 per month depending on platform selection, Otterly.ai starts from $29 per month, and several others, including Profound and Peec AI, are sales-gated with undisclosed pricing. Use these tools, and disclose which ones you used and what they cover, but do not treat any single tool's number as ground truth. We use several and we say so, rather than reselling one as an authoritative dashboard.
How honest should attribution be in GEO reporting?
As honest as the underlying data allows, which usually means less precise than a client or a stakeholder wants. Multi-channel marketing programmes produce multi-channel results, and claiming that a generative-visibility workstream alone caused a specific revenue outcome is generally not something the data supports cleanly — assisted conversions from AI exposure are directionally real but genuinely difficult to isolate from direct and branded search traffic that follows the same awareness.
The credible standard is to report what was directly measured — share of answer, citation rate, prompt coverage, citation health — as those things, and to label anything downstream, like revenue influence, as directional and caveated rather than causal. A report that presents attribution with false precision is optimizing for how it looks in a meeting, not for whether the client can trust the next one. State the limits of what you know in the same document as what you found.
What does a genuinely good GEO report actually contain?
A stated baseline, the exact prompt set used, per-engine share of answer and citation rate against that prompt set, a named competitor comparison on the same prompts, a citation-health check for dead links, and explicit caveats about sampling variance and tool coverage limits. It should also state plainly which numbers are directly measured and which are estimated or directional, rather than presenting all figures with the same false confidence.
What a good report does not contain: a single-engine snapshot presented as a full picture, a claim of causal revenue attribution without a clear methodology, a promise about future citation likelihood, or vendor tool numbers presented as independently audited when none of the available tools currently have that audit. If a report you receive from any provider is missing the prompt set it was measured against, ask for it — without it, the numbers cannot be checked or reproduced.
What does Google Search Console show about AI features, and what does it not show?
Google has begun surfacing AI-feature traffic within Search Console's existing Performance report, giving some visibility into how AI Overviews and AI Mode contribute to your Google Search performance specifically. That is a real, useful, and free data source, and it should be checked regularly.
What it does not do is give you any visibility into ChatGPT, Gemini outside Google's own surfaces, Claude, Perplexity, or Copilot — it is scoped to Google's own AI features within Search, not a cross-engine view. There is no equivalent of Search Console for the other five engines; that gap is exactly why third-party sampling tools and manually run prompt sets exist, imperfect as they are. Treat Search Console's AI-feature data as one accurate data point among several partial ones, not as the dashboard for this entire channel.
flowchart TD
A["I'm not appearing in AI answers. Why?"] --> B{"Is the page indexed and snippet-eligible in ordinary Google Search?"}
B -->|"No"| C["Fix indexation first. Nothing else matters until this is true."]
B -->|"Yes"| D{"Does robots.txt allow the retrieval agent for the engine you care about?"}
D -->|"No"| E["Unblock the specific agent, e.g. OAI-SearchBot or PerplexityBot"]
D -->|"Yes"| F{"Does the important content render as visible text, not JS-only or image/PDF-locked?"}
F -->|"No"| G["Move key facts into server-rendered HTML text"]
F -->|"Yes"| H{"Is your entity described consistently across your site, LinkedIn, Crunchbase and directories?"}
H -->|"No"| I["Unify the canonical description everywhere"]
H -->|"Yes"| J{"Does the page answer the question in the first few sentences?"}
J -->|"No"| K["Restructure to answer-first"]
J -->|"Yes"| L["You are likely visible on some engines and not others. Measure per engine with a fixed prompt set before changing anything else."]
A diagnostic order, not a checklist to complete in parallel. Fix from the top; each lower layer is unreliable to test until the ones above it are ruled out.
Risk, ethics and limits
Can you guarantee I'll be cited by ChatGPT or Google's AI Overviews?
No, and no credible party can, for reasons that are structural rather than about effort. Generated answers are non-deterministic — the same prompt produces different responses across sessions and model versions. The ranking and citation-selection logic for most engines is not publicly documented, so no one, including the vendors' own staff in many cases, can point to a formula and promise where a given lever lands. And model versions change on schedules no vendor commits to in advance, which means a result achieved today is not guaranteed to hold after the next retraining.
What can be committed to, honestly, is measurable improvement in the inputs you actually control — entity clarity, retrieval access, corroboration, answer-first content coverage — and transparent, repeated measurement of the outputs using a fixed prompt set. Any agency offering a guaranteed citation outcome is offering something the underlying systems do not support.
Can I pay to get cited by an AI engine?
Not as organic inclusion in a generated answer — that is not a purchasable inventory on any engine we are aware of, and if a vendor tells you otherwise, ask them to show you the documented mechanism, because none of the major providers publish one. What does exist, and is clearly labelled as such, are advertising products inside some AI surfaces — sponsored placements distinct from the organic synthesized answer, disclosed as ads rather than presented as neutral citations.
If a vendor offers to place your company inside ChatGPT's, Perplexity's, or Google's organic AI answers for a fee, that claim should be treated with real skepticism until they can point to a documented advertising product that matches what they are selling. Genuine visibility in the organic answer comes from the same unglamorous inputs discussed throughout this page — entity clarity, corroboration, extractable content, retrieval access — none of which are for sale as a guaranteed outcome.
What happens if a model says something factually wrong about my company?
There is no formal correction or appeal channel comparable to a legal process, and it is worth knowing that going in rather than discovering it during a crisis. If the error stems from retrieval — the model pulled from a specific, findable source — you can sometimes trace and correct the underlying source, and future retrieval-time answers may reflect the fix once that source updates and gets re-crawled. If the error is parametric, baked into training data from months or years ago, there is no direct lever to correct it quickly; it persists until a future retraining, on a schedule no vendor publishes.
The best available defense is prevention through consensus: the more consistently accurate, independent sources describe your company the same correct way, the less likely a model is to have absorbed or retrieved a wrong claim in the first place, and the more likely a wrong claim is to get outweighed over time. This is slow, and it is honest to say there is no fast fix once an error exists in a model that already reflects it.
Is any of this manipulative? Are we trying to trick AI systems?
Some tactics genuinely marketed under the GEO label are manipulative in intent, even if they do not work — fabricating reviews, stuffing keywords to game an extraction heuristic, or manufacturing a Wikipedia entry that does not meet notability standards. The KDD 2024 GEO paper itself found that keyword stuffing had a negative effect on visibility, which is a useful reminder that some manipulative tactics fail on their own terms, not just on ethical ones. Wikipedia and Wikidata communities actively remove promotional entries, and a deleted entry is a worse outcome than never having one.
What this hub argues for is a different thing entirely: making genuinely accurate information about your company easier for a machine to find, verify and quote correctly — consistent facts, clear entity data, real expertise presented clearly. That is closer to the discipline of clear, honest disclosure than to persuasion engineering. The line we try to hold is: if the tactic would look bad explained plainly to the reader it's about, don't recommend it.
What happens to my visibility when a model updates or retrains?
It can shift, sometimes significantly, without advance notice, because none of the major vendors publish a retraining or update schedule that maps to specific visibility changes. A citation pattern that held steady for months can move after a silent model version update, and there is no changelog that says "here is what changed and why your share of answer moved." This is one of the clearest ways GEO differs from SEO, where algorithm updates are at least sometimes formally announced.
The practical response is structural, not reactive: treat GEO as continuous monitoring rather than a project with an end date, exactly as the fifth phase of the Artlogic methodology frames it — baseline, re-measure on a fixed cadence, and expect some of what you observe between periods to be model-update noise rather than a result of anything you did or didn't do. Building your measurement around a fixed prompt set sampled repeatedly is what lets you tell the two apart.
What can't GEO fix?
Several real things, and naming them plainly is more useful than pretending otherwise. It cannot fix a genuinely uncompetitive product or service — no amount of entity clarity changes what happens once a buyer evaluates you against alternatives. It cannot rapidly correct false or outdated information baked into a model's parametric memory; that requires retraining on someone else's schedule. It cannot overcome a substantial body of independent, negative, or damaging coverage — consensus works against you as readily as for you, and no technique reverses well-corroborated bad information quickly. It cannot guarantee inclusion in any specific answer, because the systems themselves are non-deterministic. And it cannot substitute for basic access and indexation — no amount of downstream work matters if a page is not crawlable in the first place.
What it can do is make an accurate, differentiated, well-supported company easier to find, verify and cite correctly, consistently, over time. That is a real and valuable thing. It is not the same as control.
What does "brand safety" mean in a GEO context?
Largely the same underlying risk as reputation management always carried, applied to a new surface: the possibility that a generated answer describes your company inaccurately, unfavorably, or in a way that misrepresents your position relative to competitors, and that this happens inside a synthesized paragraph a buyer treats as a neutral summary rather than as one source's take. Because citation selection and ranking logic are largely undocumented across engines, you cannot audit exactly why an unfavorable framing occurred, only observe that it did.
Practically, this argues for monitoring mention sentiment and framing, not just raw appearance rate, as part of ongoing measurement — being named is not automatically a good outcome if the framing is wrong. It also argues for the same corroboration discipline discussed throughout this page: the more consistently accurate sources describe you correctly, the less room there is for an inaccurate framing to take hold and persist. There is no monitoring tool that catches every instance of this, which is itself worth knowing before you assume silence means nothing is wrong.
Still unanswered?
Sixty questions is a lot, and it is still not every question. If yours is not here — a specific engine behavior, a situation particular to your industry, or something about your own audit results — the honest answer is usually more useful delivered directly than guessed at in a general FAQ. Ask us.
Send us your specific question
Work with Artlogic
Artlogic builds AI visibility programmes for firms expanding across Canada, the United States, Europe and the Middle East. Google's documentation states: "As of May 7, 2026, FAQ rich results are no longer appearing in Google Search. We will be dropping the FAQ search appearance, rich result report, and support in the Rich results test in June 2026. To allow time for adjusting your API calls, support for the FAQ rich result in the Search Console API will be removed in August 2026." Shipping FAQPage markup here would not produce a Google rich result, and implying otherwise would repeat exactly the stale advice this hub argues against.
Article with BreadcrumbList is the accurate markup for what this page actually is: a single authored article organized into question-and-answer sections, described honestly to search engines rather than dressed up for a rich-result feature that no longer exists. If a future update adds FAQPage for non-Google machine consumers that still parse it, this note will be updated to say so explicitly rather than implying any Google visibility benefit.
{
"@context": "https://schema.org",
"@graph": [
{
"@type": "Article",
"@id": "https://artlogic.ca/geo/faq#article",
"headline": "GEO Frequently Asked Questions",
"description": "Sixty sourced answers on Generative Engine Optimization: how AI search actually works, what to do first, and what nobody can honestly promise you in 2026.",
"datePublished": "2026-08-06",
"dateModified": "2026-08-06",
"inLanguage": "en",
"isPartOf": { "@id": "https://artlogic.ca/geo#hub" },
"author": { "@id": "https://artlogic.ca/#organization" },
"publisher": { "@id": "https://artlogic.ca/#organization" },
"about": [
{ "@type": "Thing", "name": "Generative Engine Optimization" },
{ "@type": "Thing", "name": "AI search" },
{ "@type": "Thing", "name": "Structured data" }
],
"citation": [
{
"@type": "ScholarlyArticle",
"name": "GEO: Generative Engine Optimization",
"identifier": "arXiv:2311.09735",
"url": "https://arxiv.org/abs/2311.09735"
}
]
},
{
"@type": "BreadcrumbList",
"itemListElement": [
{ "@type": "ListItem", "position": 1, "name": "Home", "item": "https://artlogic.ca" },
{ "@type": "ListItem", "position": 2, "name": "GEO", "item": "https://artlogic.ca/geo" },
{ "@type": "ListItem", "position": 3, "name": "FAQ", "item": "https://artlogic.ca/geo/faq" }
]
}
]
}
Last reviewed 6 August 2026. Every factual claim above traces to a named source. Where the honest answer was "nobody documents this," we said so rather than filling the gap.