AI search optimization: SEO plus corroboration, facts, and measurement

AI search optimization is the work of getting a company's pages found, quoted, and recommended in answers from AI assistants. The evidence supports five parts: classic SEO, so pages enter the search indexes that assistants draw on; specific, consistent facts on those pages; mentions on independent sites; measurement that samples answers many times; and offer details that software agents can read. Most rewriting tactics sold as GEO did not hold up in controlled tests (opens in a new tab).

This guide covers how ChatGPT, Google's AI features, Gemini, Claude, Perplexity, and Microsoft Copilot choose sources, what the research shows, and the steps that follow. Facts link to their sources; recommendations are marked as ours. It does not cover ads inside assistants, or using AI tools to write SEO content, which is also sometimes called "AI SEO".

B2B buyers use these assistants to research vendors. 51% of B2B software buyers now start their research with an AI chatbot more often than with Google (opens in a new tab), according to a G2 survey of 1,076 software decision-makers and influencers in March 2026 (vendor survey).

Scope and definitions

• AI search optimization. Work that makes a company's pages retrievable, quotable, and recommendable in answers written by AI assistants. It is also called AI SEO, LLM SEO, and AI search engine optimization. AEO HQ calls its version SEO+: search engine optimization plus corroboration, facts, and measurement.
• Answer engine optimization (AEO) and generative engine optimization (GEO). Two other names for overlapping work. GEO is also the title of a 2024 peer-reviewed paper that tested nine ways of rewriting a page inside a simulated AI engine (opens in a new tab). Each approach has its own complete guide: the answer engine guide and the generative engine guide.
• AI assistant. Software that answers questions in prose and may cite web pages: ChatGPT, Google's AI Overviews and AI Mode, Gemini, Claude, Perplexity, and Microsoft Copilot.
• Retrieval. The step in which an assistant runs searches and pulls pages into the model's context before it writes. Basing the answer on those pages is grounding, and the design is called retrieval-augmented generation (RAG).
• Parametric knowledge. What a model learned in training, before any search.
• Query fan-out. Google's term for "issuing multiple related searches across subtopics and data sources (opens in a new tab)" to build one answer.
• Entity. A distinct thing, such as a company or a person, that search systems try to keep apart from others with similar names.
• Corroboration. What other sites say about a company: reviews, lists, articles, forums, and videos.
• Antipattern. A common practice that looks useful but fails when tested.

In this guide

This page is the hub for six companion pages:

1. SEO antipatterns: mistakes that also keep pages out of AI answers.
2. Technical SEO checklist for AI search: crawl, render, index, and snippet checks.
3. Entity and brand consistency checklist: one name and one set of facts everywhere.
4. llms.txt: what it is and whether it matters.
5. Brand mentions and AI recommendations.
6. How AI agents find and buy services.

The index, crawlers, and controls behind each assistant are listed in How ChatGPT, Gemini, Claude, Perplexity, and Copilot find and cite sources.

How AI search works

Two routes into an answer

An assistant can name a company from its parametric knowledge or from pages it retrieves while it answers. In a 2023 study, a model's accuracy on a factual question rose with the number of training documents about it, and retrieval reduced that dependence (opens in a new tab) (peer-reviewed; background study).

Training is the slow route for a new or small company. In controlled experiments, a model could reliably recall a fact only when its training data restated the fact in varied forms, such as paraphrases and reordered sentences (opens in a new tab) (peer-reviewed). Training data also stops at a cutoff date. Anthropic, for example, lists June 2026 as the reliable-knowledge cutoff of its newest Claude models (opens in a new tab) (official documentation).

Retrieval is the fast route, but assistants do not search every time. In a US clickstream panel, ChatGPT ran a web search on 34.5% of queries in February 2026, down from 46% in late 2024 (opens in a new tab) (vendor study). In another panel, fewer than 4% of US ChatGPT prompts about professional services carried citations (opens in a new tab) (vendor study, 2026; method not published). Claude's documentation says it searches when a request depends on current information, including "information about specific organizations, people, or products that might have changed (opens in a new tab)." Questions that name vendors or ask for current prices fit that description.

Our reading: a company that models do not yet know should work on retrieval first.

Retrieval runs on search indexes

Each major assistant documents where its retrieved pages come from. In every case the source is a search index or search provider: Google's, Bing's, Brave's, or the assistant's own.

Brave, which Claude draws on, follows Google's crawl rules: "if a domain or page is not crawlable by Googlebot, then Brave Search's bot will not crawl it either (opens in a new tab)."

The consequence is direct. A page missing from the relevant index cannot be cited through search, however well it is written. Google puts it this way: "The best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems (opens in a new tab)." Bing's guidelines say "Bing and Copilot search experiences rely on the same core crawling, indexing, and ranking foundation as traditional search (opens in a new tab)."

Google adds one condition of its own: a site must be included in the "Search generative AI" setting in Search Console (opens in a new tab). Inclusion is the default, and Google rolled the control out to all websites on August 31, 2026 (opens in a new tab).

Rank matters, for more queries than the one typed

Position inside the retrieved set strongly affects citation. In the C-SEO Bench tests, moving a page into the first positions of the model's context produced larger gains than any content rewrite (opens in a new tab) (peer-reviewed). Where AI citations overlap with ordinary search results, they skew toward the top result: 23.3% of overlapping citations were Bing's first result, and 14.5% were Google's (opens in a new tab) (preprint; 55,936 queries, July–August 2025).

Ranking for the exact prompt predicts less than it used to. In a December 2025 benchmark of more than 11,000 queries, AI Overviews and Google's organic results shared only 18% of their sources (opens in a new tab) (peer-reviewed). In another audit, 53% of the domains AI Overviews consulted were outside the organic top 10 (opens in a new tab) (preprint; 4,706 queries, 2025). The reason is query fan-out: ChatGPT "rewrites your query into one or more targeted queries" (opens in a new tab), and Google's AI features run related searches across subtopics (opens in a new tab). A page can be cited because it ranks for a sub-question the user never typed.

Our recommendation: answer each sub-question a buyer's prompt fans out into (price, scope, deliverables, timeline, comparisons, method) on a distinct page or section, and work to rank for each in Google and Bing.

Crawlers must be able to read the facts

In Vercel network data published in December 2024, none of the major AI crawlers rendered JavaScript (opens in a new tab) (vendor measurement). In an October 2025 test of a single page, only Gemini rendered JavaScript during a live fetch, and no assistant read facts that appeared only in JSON-LD (opens in a new tab) (practitioner test). Bing's guidelines warn against hiding critical content behind client-side rendering (opens in a new tab), and Microsoft advises against hiding answers in tabs, expandable menus, PDFs, or images (opens in a new tab) (official guidance).

Our recommendation: put every fact you want an assistant to repeat in server-rendered HTML.

Search crawlers and training crawlers are separate

Each company uses different crawlers for search and for training:

• OpenAI. Sites that opt out of OAI-SearchBot "will not be shown in ChatGPT search answers, though can still appear as navigational links" (opens in a new tab). GPTBot collects training data, and "each setting is independent of the others (opens in a new tab)."
• Anthropic. Blocking Claude-SearchBot "may reduce your site's visibility and accuracy in user search results" (opens in a new tab), while ClaudeBot collects training data.
• Perplexity. PerplexityBot "is not used to crawl content for AI foundation models." (opens in a new tab)
• Google. The Google-Extended robots.txt token controls whether content is used for Gemini training and for grounding in Gemini Apps and Vertex AI; it does not affect inclusion in Google Search (opens in a new tab). In one study, 21 publishers that blocked Google-Extended were never cited by Gemini (opens in a new tab) (peer-reviewed; an association, not a tested cause).

Our recommendation: allow every search crawler and the Google-Extended token. Allowing training crawlers such as GPTBot and ClaudeBot is a business decision. It is the only route into future training data, but no study has measured its effect for a single brand.

Most GEO tactics do not replicate

The best-known number in this field comes from the original GEO paper. In the original GEO experiments, adding quotations raised a source's share of the generated answer by about 41% and adding statistics by about 31%, in a simulated engine where the page was already retrieved (opens in a new tab) (peer-reviewed, KDD 2024). The outcome was the share of the answer's words credited to a page, not the chance of being cited.

Later tests with newer models and realistic pipelines did not reproduce the effect:

• A 2025 benchmark found statistically significant gains for conversational-SEO methods in 3 of 54 cases, and adding statistics lowered rank in 19 of 24 (opens in a new tab) (peer-reviewed, NeurIPS 2025). Gains also shrank toward zero as more competitors adopted a tactic (opens in a new tab).
• In a 2026 KDD study, rewriting page text to optimize it for generative engines hurt retrieval (opens in a new tab) (peer-reviewed). In a test collection of 171,003 real web pages, an automated rewriting method (AutoGEO) lost 22.35 retrieval ranks, because it made pages longer and diluted them (opens in a new tab).
• A review of 45 studies rejected "GEO increases visibility by 40%" as a general claim, and found no technique with a stable, long-term causal effect on discoverability across platforms (opens in a new tab) (preprint).
• A detector flagged GEO-rewritten pages with an F1 score of 0.944 out of 1.0, and estimated that 8.9% of 10,095 pages in real Google and Gemini results were GEO-optimized (opens in a new tab) (preprint).

Other popular tactics show little or no measured benefit:

• Formatting alone. In 252,000 controlled trials across six models, structured versus dense text had no consistent effect on which source a model cited first (opens in a new tab) (peer-reviewed, lab setting; the authors work for a marketing-software company). Structure may help earlier: better titles, meta descriptions, headings, and schema fields raised the retrieval hit rate by 22% (opens in a new tab) in the KDD pipeline.
• Structured data. Across 1,885 pages that added JSON-LD markup, AI citations changed by −4.6% in AI Overviews (a small, significant decline), and by +2.4% in AI Mode and +2.2% in ChatGPT, both statistically indistinguishable from zero (opens in a new tab) (vendor study; every page studied was already cited). Google says "there's no special schema.org markup you need to add (opens in a new tab)."
• FAQ markup. Google stopped showing FAQ rich results on May 7, 2026 (opens in a new tab).
• llms.txt files. Of about 38,000 domains that published a valid file, 97% received no requests for it in May 2026 (opens in a new tab) (vendor study, server logs), and Google Search, including its generative AI features, does not use AI text files (opens in a new tab).

Our recommendation: do not rewrite pages to fit what engines are thought to prefer. Rewriting that adds no information has no reliable benefit and can reduce retrieval. AEO HQ's review of which findings replicate covers these studies in more depth.

What does replicate

The findings that hold across studies concern relevance, where the answer sits, and concrete facts.

• Match the question. In an ACL 2024 study, relevance to the question predicted which evidence four of five models found convincing, while scientific references or a neutral tone had "neutral to negative" effects (opens in a new tab) (peer-reviewed). In the six-model study, an on-topic source was cited first far more often than an off-topic one (odds ratios of 221 to over 10,000) (opens in a new tab), and missing the query's own terms was a significant disadvantage (opens in a new tab) (peer-reviewed, lab setting). An odds ratio of 221 means the odds were 221 times higher.
• Put the answer first. Rerankers scored pages higher when the answer appeared early, and moving it later caused "significant rank drops" (opens in a new tab) (peer-reviewed). The only controlled field study found a 1.82-fold rise in ChatGPT referrals (95% confidence interval 1.31–2.54) after changes that included question-form titles and two-to-three-sentence standalone summaries (opens in a new tab), but a placebo test returned p = 0.16, so the authors call it suggestive (preprint; one site, owned by the authors' employer).
• State decision facts. Explicit prices were a significant citation driver in all six models (odds ratios 6.3 to over 10,000), as was a 2026 rather than a 2019 date (14.4 to over 10,000); specifications, comparisons, evidence, confident wording, and consistent claims had smaller effects, significant in at least four of six models (2.1 to 243) (opens in a new tab) (peer-reviewed, lab setting).
• Show evidence the model can see. In lab tests, a well-known brand won every choice when products were equal, but a fictional competitor won half the time with a 0.075-star rating edge, 1.6 times the reviews, or a 7.3% lower price (opens in a new tab) (preprint).
• Keep content current, honestly. Seven LLM rerankers promoted passages given newer dates, shifting the top 10 by up to 4.78 years (opens in a new tab) (peer-reviewed). Across about 17 million citations, AI assistants cited pages about 25% "fresher" than Google's organic results, while Google's AI Overviews did not favor newer pages (opens in a new tab) (vendor study).

Buyers check the same facts. In a January 2026 survey of 1,862 technology buyers, 94% of those who use AI said they fact-check its output at least some of the time (opens in a new tab) (vendor survey). In a March–April 2026 survey of US B2B professionals, 71% of those who use AI at work said they visit a vendor's site after an AI mentions it (opens in a new tab), and 27% of them said AI vendor recommendations don't reflect real pricing or contract structures (opens in a new tab) (vendor survey; 519 respondents who use AI at work).

Our recommendation: publish the facts buyers ask assistants about, such as price, what is included, turnaround, who does the work, and terms, as plain text. Keep them identical everywhere, and change a page's date only when its content changes.

Third-party corroboration

AI answers lean heavily on what other sites say about a company:

• ChatGPT drew 93.5–95.1% of the domains it cited from earned media, such as reviews, comparisons, and editorial coverage, while Google's results drew more on brand-owned and social pages (opens in a new tab) (preprint; ranking-style prompts sent through APIs, mid-2025).
• Of 149,912 citations from five engines, 2.9% pointed to the tracked brand's own domain, and about 21% to "best-of" listicles (opens in a new tab) (preprint, March–May 2026; the author co-founded a visibility-tracking company).
• Across 75,000 brands, branded web mentions correlated at 0.66 to 0.71 with how often a brand appeared in ChatGPT, AI Mode, and AI Overviews, and YouTube mentions at about 0.74, while backlinks showed "very weak correlations" (opens in a new tab) (vendor study).
• 112 recently launched startups surfaced in only 3.3% of discovery queries to gpt-4o-mini without web access and 8.3% to Perplexity; on Perplexity, referring domains (r = 0.319) and Reddit presence (r = 0.395) predicted discovery, and an on-page GEO score did not (opens in a new tab) (preprint, thesis-based).
• On a first answer to a generic prompt, household brands appeared 73% of the time, mid-market brands 44%, and small brands 11% (opens in a new tab) (preprint; 102 brands).

These studies are correlational: a brand's size can drive both its mentions and its visibility. They show where assistants look, not that one new mention changes an answer. Google also says "seeking inauthentic 'mentions' across the web isn't as helpful as it might seem (opens in a new tab)."

Names need to be unambiguous. Language models "often yield ambiguous answers or incorrectly merge information belonging to different entities" (opens in a new tab) that share a name (peer-reviewed). Google's article markup documentation asks for an author URL that "uniquely identifies the author" and accepts links to the author's other profiles (opens in a new tab), marked up as sameAs. No study has shown that such markup raises AI recommendations.

Our recommendation: earn mentions where assistants look, through third-party reviews, comparisons that publish their method, original data that others cite, and disclosed participation in buyer communities. Never pay for placement or post anonymously about your own company.

Measurement: answers are distributions

AI visibility cannot be read from one test, because answers change from run to run:

• When the same prompt was repeated, ChatGPT and Google's AI returned the same list of brands less than once in 100 runs, and Claude only slightly more often (2,961 runs, November–December 2025) (opens in a new tab) (vendor study; a co-investigator works for a visibility-tracking vendor).
• Cited sources overlapped only 32–43% across repeated runs, and about 65% changed from one day to the next (opens in a new tab) (preprint; German-language prompts, January–March 2026; the first author is affiliated with a measurement vendor).
• The app and the API of the same assistant shared only 12–15% of cited domains (opens in a new tab) (preprint, September 2026).
• Rates over many runs are steadier: 77.5% of brand, prompt, and engine combinations were either always or never mentioned (opens in a new tab) (preprint).

Our recommendations, based on these studies:

1. Report the share of runs that mention the company, per assistant and prompt set, never a "rank in ChatGPT". Share of voice against fixed competitors is a useful second metric.
2. Run each prompt several times. At least 7 to 8 runs per prompt brought the standard error of a per-prompt detection rate below 0.10, and 3–4-week rolling windows reduced it further (opens in a new tab) (preprint). Standard error estimates how far a measured rate may be from the true rate.
3. Add prompts before runs, because repeated runs reduce only the variation within a prompt (opens in a new tab) (preprint). Use Wilson score or Bayesian intervals, because normal-approximation intervals are too narrow below a few hundred data points (opens in a new tab) (peer-reviewed, ICML 2025).
4. Test the consumer apps, not only APIs, and log the date, product, and mode.

First-party data covers part of this. Search Console's Generative AI performance report shows impressions, not clicks, in AI Overviews and AI Mode (opens in a new tab). Bing Webmaster Tools' AI Performance report shows citations, cited pages, and grounding queries for Copilot and Bing's AI summaries (opens in a new tab). Google Analytics 4's AI Assistant channel counts visits from assistants such as ChatGPT, Gemini, and Copilot, but counts AI Overviews and AI Mode as Organic Search (opens in a new tab). ChatGPT adds utm_source=chatgpt.com to referral URLs (opens in a new tab).

AI referral traffic is small: in a February 2025 study of 3,000 sites, 0.17% of visitors came from AI assistants (opens in a new tab) (vendor study). Google users clicked a traditional search result on 8% of visits when an AI summary appeared, compared with 15% when none appeared (opens in a new tab) (Pew Research Center; 900 US adults, March 2025). Whether AI visitors convert better is contested: across 973 e-commerce sites (Aug 2024–Jul 2025), ChatGPT referrals converted below organic search, paid search, email, and affiliate traffic, and above paid social only (opens in a new tab) (peer-reviewed; the data came from the first author's employer).

Being cited is not the same as being described correctly: 11.0% of 98,020 claims in AI Overviews were not supported by the pages they cited (opens in a new tab) (preprint, March–April 2026). Check what assistants say about your company, not only whether they cite it.

No study has measured how long a new company takes to be recommended; a 2026 review of 45 studies found no technique with a stable, long-term, cross-platform causal effect on discoverability (opens in a new tab) (preprint). In the best observational data, the fastest-moving small brands gained 10 to 20 points of visibility between March and May 2026 (opens in a new tab) (preprint).

Agent-readable offers

Software agents now research and sometimes buy on a person's behalf. As of September 2026, no major assistant sells a professional service inside the chat. On March 24, 2026, OpenAI stopped offering in-chat Instant Checkout to new merchants; purchases that start in ChatGPT are completed on merchants' own sites (opens in a new tab). Google's checkout in AI Mode and Gemini lists "Services: lessons, online classes, and travel packages" among restricted products (opens in a new tab), and Anthropic's software directory excludes software that "executes financial transactions on behalf of users" (opens in a new tab).

Agents can still buy on a seller's own website, with a person approving the payment. ChatGPT's cloud browser asks for confirmation before payments (opens in a new tab). Stripe documents how a seller can accept service bookings and payments from agents (opens in a new tab), and in its Link agent wallet the buyer approves each spend, with payers limited to US and Canadian consumers (opens in a new tab) (official documentation; preview features).

What sways agents is still being studied, mostly in simulations. Ratings were the only promotional cue that consistently raised selection across models (opens in a new tab) (working paper). Perplexity says its shopping feature is "more likely to recommend merchants who provide deeper product details such as availability, reviews, pricing, and specifications (opens in a new tab)."

Our recommendation: publish the full offer in plain HTML (price, currency, scope, deliverables, turnaround, refund terms, and who does the work), allow checkout without a login or a sales call, keep verified agents out of firewall blocks, and never add hidden instructions for AI. The guide to how AI agents find and buy services covers protocols and checkout.

Steps

The steps are our recommendations; the facts in them link to sources.

1. Check crawl access. robots.txt should allow Googlebot, Bingbot, OAI-SearchBot, Claude-SearchBot, and PerplexityBot, and should not disallow Google-Extended. Check the host's firewall too: OpenAI requires hosts and CDNs to allow traffic from its published searchbot IP addresses (opens in a new tab).
2. Get indexed by Google. Submit a sitemap and request indexing once per key URL. Crawling "can take anywhere from a few days to a few weeks," (opens in a new tab) and Google uses a sitemap's lastmod date only when it is "consistently and verifiably" accurate (opens in a new tab).
3. Get indexed by Bing. Verify the site in Bing Webmaster Tools and send new URLs through IndexNow. IndexNow shares each submission with Bing, Yandex, Naver, Seznam.cz, Yep, and Amazon, but not Google, and "Submitting a URL does not guarantee immediate indexing." (opens in a new tab) In 2023, Bing said 12% of new URLs clicked in its results were first discovered through IndexNow (opens in a new tab).
4. Check Brave. Search Brave for your brand and key pages. Brave documents its submit-URL form only for re-fetching a page after a noindex directive is added and for delisting pages that no longer exist (opens in a new tab); nothing in its documentation says the form adds new pages or speeds up their discovery.
5. Put the facts in server-rendered HTML, not only in widgets, tabs, PDFs, or JSON-LD.
6. Map the questions buyers ask and their sub-questions. Give each distinct intent one page, with the buyer's wording in the title and a direct answer first.
7. Make the facts consistent across your site, profiles, and listings.
8. Earn corroboration through reviews, method-disclosed comparisons, original data, and named community participation.
9. Make the offer readable and buyable by people and agents.
10. Measure with a repeated prompt panel, plus Search Console, Bing Webmaster Tools, and analytics.
11. Update honestly. Change dates only when content changes.

What the evidence shows and does not show

Antipatterns

Each of these looks useful and fails for a documented reason. The full SEO list has more, with a detection test for each.

Checklist

The technical SEO checklist and the entity checklist hold the full versions. The core checks:

Frequently asked questions

Is AI search optimization different from SEO?

Partly. Assistants retrieve pages from search indexes, so the base is SEO; Google says "The best practices for SEO continue to be relevant (opens in a new tab)" for its AI features. The additions are specific facts, corroboration on other sites, measurement across assistants, and agent-readable offers.

Do strong Google rankings still influence AI citations in 2026?

Yes, but not only for the query as typed. Overlapping citations skew toward the top-ranked result (opens in a new tab), yet AI Overviews and organic results shared only 18% of their sources (opens in a new tab), because assistants split questions into sub-queries. Bing rankings matter too, because Copilot searches Bing (opens in a new tab).

Can anyone guarantee that ChatGPT will recommend my company?

No. OpenAI says "Placement is not guaranteed." (opens in a new tab) Google says third-party tools "can't guarantee performance," (opens in a new tab) and Bing's guidelines say "GEO does not guarantee grounding or citations." (opens in a new tab)

Do we need llms.txt or schema markup?

Current evidence shows neither is needed to be cited. Most llms.txt files are never requested (opens in a new tab), though Stripe's directory for AI agents asks sellers for a link to theirs (opens in a new tab). Adding schema did not raise AI citations in a matched study (opens in a new tab), but Google still recommends structured data for its rich results (opens in a new tab).

What do LLM SEO checkers measure, and can I trust one?

An LLM SEO checker asks AI assistants a set of prompts and reports whether a brand appears. Because the same prompt rarely returns the same brand list (opens in a new tab), one run is one sample. Trust a checker only if it states its prompts, runs, assistants, and dates, and reports intervals.

Why does ChatGPT recommend my company to me but not to my customers?

Answers vary with the wording, the surface, the person asking, and from run to run: revealing a user's identity changed chatbot recommendations (opens in a new tab) (peer-reviewed). The effect of an assistant's memory of past chats has not been studied. A founder's own test is one sample; a prompt panel run many times gives the rate.

How long does AI search optimization take?

No study has measured it. Indexing alone can take a few days to a few weeks in Google (opens in a new tab). Our expectation, which is not a forecast: citations on specific, fact-based prompts come before mentions on broad "best vendor" prompts, which depend on corroboration built over time.

Next steps

To check your own site against the retrieval and fact items in this guide, AEO HQ sells a fixed-price Instant AEO Audit for $499. It includes a live crawl of your key pages, AI crawler access checks, and a side-by-side analysis of up to three competitors.

Change log

• September 27, 2026: First published.