meo / auditRequest a walkthrough

Meo audit guides · Meo Advisors

AI Search Visibility: Mentions, Citations, and Access

Audit how your business appears in AI answers. Separate mentions from citations and search access from training-crawler preferences.

Published · Reviewed

Overview

An AI search visibility audit tests whether your business is mentioned or cited in a defined sample of answers and checks public access signals. Results depend on the questions, engine, and collection conditions. Search and answer access should be reviewed separately from permissions for collecting content used in model training.

Define the question before counting the answer

An AI assistant can answer a research question, compare named providers, or suggest businesses for a service. Those are different tests. Start by deciding which buyer task you want to examine and how the question represents it.

A neutral discovery prompt might ask for commercial cleaning providers in a defined city. A comparison prompt names the alternatives. A branded trust prompt asks about a particular business. Mixing these answers into a single visibility rate makes the result hard to interpret.

Keep the question wording, location, language, engine, product surface, and date. Record whether the observation came from an API, a captured search interface, or another defined instrument. Do not assume those surfaces produce identical answers.

Count mentions and citations separately

A mention means the answer names the business. A citation means it links to a supporting source. A business can appear without its website being cited, and its website can be cited as an information source without the business being recommended.

Check entity identity. Similar business names, old domains, and directory listings can complicate extraction. Confirm the domain and market before counting a result. Preserve ambiguous matches for review rather than silently accepting them.

Recommendation wording also needs context. A business listed among options is different from the answer explicitly endorsing it. Keep descriptive notes close to the sampled answer and avoid treating an apparent list position as a universal ranking.

For source review, inspect which public pages support the answer. Are they current and relevant to the service? A citation is evidence of what the answer linked to; it does not establish that the source is correct or that every generated statement follows from it.

Separate training controls from search and answer access

Crawler names can look similar while serving different purposes. Review them against the current vendor documentation. Grouping every bot under "allow AI" can change permissions you did not intend to change.

Training and collection controls

Training and collection controls
TokenDocumented role and interpretation
GPTBotCollects material that may be used for OpenAI model training; separate from ChatGPT search. OpenAI crawler docs
ClaudeBotCollects web content that may contribute to model training. Anthropic crawler docs
Google-ExtendedA product-control token covering future Gemini training and certain grounding uses, without controlling inclusion or ranking in Google Search. It is not a separate HTTP crawler identity. Google crawler docs
CCBotBuilds Common Crawl's open web dataset. Treat it as collection access relevant to downstream training decisions, not a live answer-search crawler. Common Crawl FAQ
Sources: [1] overview-of-openai-crawlers[2] does-anthropic-crawl-data-from-the-web[3] google-s-common-crawlers[4] common-crawl-faq

The "training" shorthand therefore needs qualifications. Common Crawl is a general web dataset, and Google-Extended covers more than training. Neither should be treated as the access switch for ChatGPT search or Google Search answers.

Search crawlers and user-directed retrieval

Search crawlers and user-directed retrieval
AgentDocumented role and interpretation
OAI-SearchBotSupports inclusion in ChatGPT search results. OpenAI crawler docs
ChatGPT-UserFetches content for certain user actions. It does not control search inclusion, and robots rules may not apply to these user-directed requests. OpenAI crawler docs
PerplexityBotSupports Perplexity search results; the vendor distinguishes it from foundation-model training. Perplexity crawler docs
Claude-SearchBotSupports search-result quality and relevance. Anthropic crawler docs
GooglebotCrawls for Google Search; relevant Search access and indexing requirements apply to AI Overviews and AI Mode. Google AI feature docs
Sources: [1] overview-of-openai-crawlers[2] perplexity-crawlers[3] does-anthropic-crawl-data-from-the-web[4] ai-features-and-your-website

User-directed retrieval also includes Perplexity-User and Claude-User. Perplexity says its user fetcher generally ignores robots rules because a user requested the fetch. Anthropic describes its user agent separately from search and training. Perplexity docs, Anthropic docs. Sources: [1] perplexity-crawlers[2] does-anthropic-crawl-data-from-the-web

Review robots policy alongside hosting and firewall behavior. A public declaration of permission does not establish that a genuine request succeeds. Likewise, a user-agent string alone does not authenticate the sender. Follow the vendor's verification guidance when your technical team inspects access logs.

What to check

Build a reproducible prompt and access record:

  • Confirm the business, domain, services, and relevant market.
  • Separate neutral discovery, comparison, and branded reputation prompts.
  • Save exact wording and the engine or product surface.
  • Record collection date, language, and location context.
  • Count completed answers separately from unavailable requests.
  • Distinguish a business mention from a domain citation.
  • Check ambiguous entity matches before counting them.
  • Preserve the cited source and its relevance to the answer.
  • Review search and retrieval access separately from training preferences.
  • Check robots rules and relevant infrastructure restrictions.
  • Review public service explanations and consistent business identity.
  • Label conclusions from a small answer panel as directional.

If the intended buyer question is wrong, repeating it more often will not repair the study. Review prompt relevance with the service owner before interpreting a result.

Show the denominator behind every rate

Mention rate needs a count of completed answers and a definition of what qualifies as a mention. Citation rate needs its own definition. Keep results separated by engine before combining anything.

An unavailable answer is missing evidence. It should not count as an answer that omitted the business. If the denominator changes between reviews, report that change before presenting an apparent improvement.

Treat answer variation as part of the method

Wording, date, engine, and retrieval context can change the answer. Repeats can help reveal variability, but they still describe the chosen panel. Record the method so another review can use comparable conditions.

A single omission cannot establish that the business is absent from AI search. A favorable answer also cannot establish durable leadership. Look for patterns that justify a specific content or access investigation.

Keep reputation interpretation separate. A generated summary may contain errors or mix sources. Verify a concerning statement against its cited public material before treating it as a business finding.

Access is permission, not a recommendation guarantee

Allowing a search crawler removes a possible access obstacle. It does not guarantee that a system crawls, indexes, cites, or recommends your page. A content correction likewise needs later observation rather than an assumed visibility gain.

Google says its AI Search features do not require special AI files or special schema. Existing eligibility and search fundamentals remain relevant. Source: Google AI feature guidance. Treat claims about mandatory proprietary optimization files cautiously. Sources: [1] ai-features-and-your-website

Improve the public material an answer could use

Start with clear, supportable explanations of your service, audience, and coverage. Answer a real buyer question directly, then explain the conditions and limitations. Keep the business identity consistent between the page and its structured information.

When an answer cites outdated information, identify the actual source before editing your homepage. The correction may belong on an old service page, public profile, or third-party listing. Your team can request an update where it has a valid basis to do so.

Prioritize a verified access defect or factual inconsistency ahead of speculative copy changes. Record the baseline, correction, and follow-up method. Avoid claiming causation from an answer that happened to change afterward.

Common mistakes: blocking training and assuming search is blocked too

The controls differ by vendor and purpose. Review the exact agent or token. A policy choice about training should not automatically be classified as a search-access failure.

Common mistakes: asking the engine to include your brand

That can test recognition or compare named options, but it is not neutral discovery. Keep branded prompts in their own panel so the report does not inflate spontaneous visibility.

Common mistakes: reporting "AI market share" from a custom panel

A panel share depends on selected questions, engines, and competitors. Label it as a derived sample comparison. It does not measure total market demand or all AI-user exposure.

Common mistakes: adding unsupported claims for citation appeal

Invented outcomes, superlatives, and fabricated proof weaken the information a buyer can verify. Publish concrete service facts and explain uncertainty. Do not manufacture authority signals to influence an answer.

Keep the instruments and denominators separate

Meo labels ChatGPT collected through the DataForSEO UI scraper separately from Perplexity sonar API answers and Google AI Overviews in search results. They are different instruments and product surfaces; their answer sets must not be presented as interchangeable or combined into a universal assistant ranking. Sources: [1] chatgpt-ui-scraper[2] sonar-api

For each engine, the discovery numerator counts qualifying business mentions; its denominator counts completed neutral discovery cells. Show unavailable cells separately as missing. A citation rate counts own-domain citations among those same completed cells. A named AND cited measure requires both a matched business name and an own-domain citation in the same answer; neither alone satisfies it.

Match confirmed aliases, official domains, and the market before counting a mention. Save the exact prompt, prompt version, engine, instrument, date, language, and location. Comparison and reputation prompts stay separate because naming the brand in a question changes what an answer can establish.

The deep manifest plans up to 24 prompts: 12 discovery, 6 comparison, and 6 reputation, across two engines with two repeats. Those engines are the ChatGPT UI scraper and Perplexity sonar API. The funded manifest and completion record govern delivery; unavailable cells do not become completed observations.

The instant discovery sample is illustrative. Mentions, citations, and relative discovery checks can earn partial or pass, but sample misses cannot alone produce an F or a negative reputation claim. Verified access and identity defects remain separate evidence. Repeated panels can describe variation without establishing why it occurred.

How Meo reads crawler directives

Meo applies bot-policy/v1 to the intended public path. The graded search and answer agents are OAI-SearchBot, ChatGPT-User, PerplexityBot, Perplexity-User, Claude-SearchBot, Claude-User, and Googlebot. Training or collection tokens GPTBot, ClaudeBot, Google-Extended, CCBot, and Applebot-Extended are observed only, with no penalty for restricting them.

Select the longest matching user-agent group and merge equal-specific groups. Use the wildcard group only when no specific group matches. Within the selected rules, the longest matching path wins; Allow wins an equal-length Allow/Disallow tie. A broad wildcard block does not override a more specific permitted search-agent group. Sources: [1] robots-spec

These are declared policy observations, not proof that a product actually retrieves or includes the page. OpenAI and Perplexity describe exceptions for user-directed fetching, so a robots restriction for a user agent does not prove a live fetch was blocked. Review hosting and firewall evidence separately before diagnosing access. Sources: [1] overview-of-openai-crawlers[2] perplexity-crawlers

How the free Meo audit checks this

The free Meo audit reviews public crawler directives, machine-readable identity, readable service material, and a small directional answer panel. It labels engine, prompt version, date, and numerator/denominator per instrument.

Discovery, comparison, and reputation stay separate. Missing answers are not misses, training restrictions carry no penalty, and illustrative omissions do not justify a failing AI discovery verdict. A mention and a citation are different observations, neither a promise of future recommendation.

The full report adds the funded discovery/comparison/reputation panel and repeats, source review, and competitor context where available. It offers supported follow-up questions without assuming a later answer change was caused by an edit. The full report is free with email and phone; automated calls and texts are optional.

Read the audit methodology for the scoring rules, coverage requirements, source policy, and limits.

Frequently asked questions

Does ChatGPT-User control search inclusion?

No. OpenAI describes it as a user-directed agent and identifies OAI-SearchBot as the control for automated search crawling. An access audit should preserve that distinction instead of treating the agents as interchangeable. Sources: [1] overview-of-openai-crawlers

Is Google-Extended the AI Overviews switch?

No. Google's documentation separates Google-Extended's uses from Google Search inclusion and ranking. Review Googlebot access and Search eligibility for AI Overviews and AI Mode, while assessing Gemini training and grounding preferences separately. Sources: [1] google-s-common-crawlers

Do you need an llms.txt file?

Do not treat it as a universal requirement or a recommendation guarantee. Google explicitly says special AI text files are unnecessary for its Search AI features. Focus first on accessible, accurate public content and the controls documented by each vendor. Sources: [1] ai-features-and-your-website

Can the audit guarantee future citations?

No. It can document current observations, identify supported access or content issues, and define a repeatable follow-up. Future answers remain dependent on the engine, question, sources, and collection conditions.

Sources and review dates