# AI Visibility Audit: 7 Checks Before Hiring an AI SEO Agency **URL:** https://mettevo.com/blog/article/ai-visibility-audit **Published:** 2026-10-26 **Updated:** 2026-10-26 **Author:** Oleg Silin **Category:** Blog | Mettevo > Discover how your brand appears in AI search with an AI Visibility Audit from Mettevo Digital Agency. Identify gaps and improve visibility. ![AI Visibility Audit](https://stage.mettevo.com/wp-content/uploads/2026/10/Professionals_reviewing_AI_SEO_a…_20261026093658.jpg) --- An artificial intelligence (AI) visibility audit is a test of whether AI search products can reach, understand and cite your pages. Run it before hiring, because seven checks need a browser, a terminal and Search Console, and they turn an agency pitch into claims you can verify. [Google says](https://developers.google.com/search/docs/appearance/ai-features) AI Overviews need no special optimization. Google's [generative AI guide](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide), updated July 10, 2026, adds that structured data is not required and that Google Search does not use llms.txt. So an honest AI visibility audit finds closed doors, conflicting facts and missing evidence, not a secret tactic. Ordinary [search engine optimization](https://mettevo.com/seo) (SEO) fundamentals still come first. We sell this service, so we ran or attempted the seven checks on mettevo.com and report where we passed, failed or could not finish. ## **Key Takeaways** - Google says AI Overviews and AI Mode need no special optimization, so an AI visibility audit tests access, clarity and evidence (Google Search Central, December 10, 2025). - Answers drift: Ahrefs found a 70% chance that a Google AI Overview changes between observations. - On [mettevo.com](http://mettevo.com), our September 28, 2026 crawl found none of the 12 AI user agents blocked and structured data on all 58 addresses that returned status 200. We cannot pass Checks 4 and 5: Ahrefs Brand Radar had no configured reports for our nine properties. ## **What are the 7 checks in an AI visibility audit?** Checks 1 and 6 test the same door from two sides: the rules a site publishes and the responses its server returns. The pass rules are our judgment, not a vendor standard. Check Pass Fail (red flag) 1\. robots.txt and llms.txt No search or user-fetch agent is blocked by accident. A blanket Disallow or an unexplained block. 2\. Schema.org markup per template Each template has the types it needs. A template with none, or markup for hidden text. 3\. Entity profiles Every profile that exists agrees on the facts. Different names, dates or descriptions. 4\. ChatGPT and Perplexity queries A dated log with accurate, citing answers and one category mention in 15 runs per engine. No log, wrong facts or no mention. 5\. AI Overviews, top 20 queries A dated log of all 20 queries names the cited domains. No log, or you rank top 10 and others are cited. 6\. Crawlability for named AI agents Same status as a normal browser, usually 200. Status 403, 429 or a challenge page. 7\. Content format The opening answers the target query. The answer sits under an introduction. ### **1\. Do robots.txt and llms.txt let the 12 AI agents in?** Open yourdomain.com/robots.txt and yourdomain.com/llms.txt. For each agent below, find the User-agent group that names it. [Google's documentation](https://developers.google.com/search/docs/crawling-indexing/robots/robots_txt) says the \* group applies when none does. Look for Disallow lines that cover content, such as a staging leftover. Type Agents Documented role Search index OAI-SearchBot, PerplexityBot Surfaces sites in ChatGPT and Perplexity search. User-triggered fetch ChatGPT-User, Perplexity-User, Claude-User Fetches a page when a user asks. Model training GPTBot, ClaudeBot, meta-externalagent Training data. Meta adds product indexing. Archive and other CCBot, Bytespider Common Crawl's open archive. No vendor page found for Bytespider. Control tokens Google-Extended, Applebot-Extended Opt-outs from Gemini and Apple model use. No crawling of their own. Roles follow OpenAI, Perplexity, Anthropic, Google, Apple, Meta and Common Crawl. OpenAI says sites that opt out of its search crawler will not be shown in ChatGPT search answers, while disallowing GPTBot only signals that content should not train its models. Robots.txt rules may not apply to ChatGPT-User, and Perplexity-User generally ignores them. llms.txt is a Markdown file proposed on September 3, 2024. Google's guide says Google Search does not use such files, and maintaining one is fine. In [Ahrefs' server-log study](https://ahrefs.com/blog/llmstxt-study/) of 137,210 domains (June 15, 2026), 97% of about 38,000 llms.txt files received zero traffic in May 2026. Publish it because it is cheap, not because it is proven. In our September 28, 2026 crawl, none of the 12 agents was blocked and llms.txt was present. Re-reading robots.txt on October 22, 2026, we found 16 agents named in one group with an explicit Allow rule, including 11 of our 12. Claude-User is not named, and the default group does not block it. Claude-SearchBot, Anthropic's search agent, is on neither our list nor the file. We have not read server logs. ### **2\. Does every page template carry schema.org markup?** List your page templates. Read the schema.org structured data in the source of one page per template and compare it with the visible text. Google's guide lists overfocusing on structured data among the things to ignore. Do not score FAQPage as a win: Google says rich results for frequently asked questions (FAQ) stopped appearing from May 7, 2026. The September 28, 2026 crawl found no page without structured data. The home page and /about-us carry Organization, the blog index carries WebSite but no Organization, and the 54 articles carry BlogPosting, Person and Organization. Fifteen articles add FAQPage. The crawl counts types, not validity, and it never reached service pages such as /seo. ### **3\. Do your entity profiles agree with each other?** An entity is the company as machines identify it. Search your brand name for a knowledge panel: [Google says](https://support.google.com/knowledgepanel/answer/7534842) panel information generates automatically from public information on the web. Then open Wikidata, Crunchbase and your LinkedIn company page and compare name, website, founding year and description. Wikidata accepts items with serious public references, so no item is normal for a small company. List profiles in Organization markup with [sameAs](https://developers.google.com/search/docs/appearance/structured-data/organization). This check audits facts. It does not buy mentions: Google's guide lists seeking inauthentic mentions among the things to ignore. On October 22, 2026, a Wikidata search for Mettevo in English returned one item, a farm in Norway, so we have no item to reconcile. Our About page links a LinkedIn company page, which we did not open; the page links no Crunchbase profile. Crunchbase refused our request with status 403, and we did not test for a knowledge panel. We do not score ourselves on Check 3. ### **4\. Do ChatGPT and Perplexity name and cite your brand?** Write five branded queries, such as your company name plus reviews, and five category queries a buyer would type. Run each three times in ChatGPT with web search on and in Perplexity. Log the date, engine, whether the brand is named, whether the facts are right and which domains are cited. [OpenAI says](https://developers.openai.com/api/docs/bots) sites that opt out of its search crawler will not be shown in ChatGPT search answers, so a Check 1 block appears here as an absence. Ahrefs measured run-to-run change only for Google AI Overviews, so the three-run rule is our judgment. For mettevo.com we have no dated citation log to publish. Ahrefs Brand Radar returned no configured reports for any of our nine properties, and Search Console data pulled through Ahrefs has no AI surface. By the table, Check 4 is a fail for us. ### **5\. Do you appear in Google AI Overviews for your top 20 queries?** Export your top 20 queries by clicks from Search Console (Performance report, Web search type, last three months). Search each in a private window from your target country and log whether an AI Overview appears and whether your domain is among its links. Then open the [Generative AI performance report](https://support.google.com/webmasters/answer/16984139), which shows impressions in AI features by page, country, device and date. It counts impressions only. ### **6\. Can the 12 AI user agents fetch your pages?** Check 6 tests the responses, because a firewall or content delivery network can refuse a bot that robots.txt welcomes. Ten of the 12 agents send their own requests. Google-Extended and Applebot-Extended do not: Google says the first has no separate user agent string, and Apple says the second does not crawl webpages. Request your home page and one article as each of the other ten: for ua in OAI-SearchBot GPTBot ChatGPT-User PerplexityBot Perplexity-User ClaudeBot Claude-User CCBot Bytespider meta-externalagent; do   printf '%s ' "$ua"; curl -s -o /dev/null -w '%{http\_code}\\n' -A "$ua" https://example.com/ done A laptop request tests user-agent rules, not network-address checks, so allow-list from the address ranges that OpenAI and Perplexity publish and confirm in server logs. Also fetch the page with curl and search it for your main heading. If the text is missing, JavaScript adds it. Our [guide to headless content management systems for Next.js](https://mettevo.com/blog/article/best-headless-cms-for-next-js-in-2026-complete-comparison-integration-guide) explains the rendering modes. Check 6 is open for mettevo.com: we did not run the ten-agent request test or read server logs. With JavaScript rendering off, our crawler still read titles, main headings and structured data on 58 addresses, with a median of 2,190 words per page, so the content is in the server response. ### **7\. Do your pages match the shape of the questions people ask?** Export the Performance report by page for two equal windows. Tag each page by the query it answers: a question-shaped page opens with the answer to one query, and a listicle ranks several options. Google's guide says you need not write in a specific way just for generative AI search, so treat any split as a hypothesis. On [mettevo.com](http://mettevo.com) (domain property, Pages tab, Web search type, September 22 to 28, 2026 against September 15 to 21), our explainer on invalid web addresses, which opens with a definition, went from 16 to 23 clicks, up 44%. Impressions stayed flat (3,814 to 3,852) and average position moved from 6.28 to 5.65. Two roundups, our free Ahrefs alternatives list and our Firefox plugins list, had zero clicks in both weeks. Their impressions fell 67% (4,377 to 1,458) and 5% (614 to 584). The first list's average position improved from 18.46 to 16.22 even as impressions fell, and we have not diagnosed why. CTR of about 0.6% on the invalid-URL page at position 5.65 is also well under what the position typically earns on blue links, which can reflect AI Overview coverage, a mix of lower-ranking queries behind the average, or a high share of unbranded informational traffic. We have not isolated the cause here. This is not format evidence. Seven extra clicks are a small difference, the explainer's position also improved, and Google's [September 2026 spam update](https://status.search.google.com/incidents/XhUDXP7A67iHCD2kmbVu) began on September 24. Our Core Web Vitals guide, which is not a listicle, fell from 4,117 to 81 impressions that week. ## **What can an AI visibility audit not tell you?** A self-audit shows whether the doors are open, whether your facts agree and what a dated snapshot looks like. It cannot show four things: - Share of voice and stability. Ten queries on two engines are a sample, and one run proves presence on that run, never absence. - Cause. An open door does not explain a citation, because Google says no special optimization is required. - Business impact. The Generative AI performance report counts impressions only, so clicks and leads need analytics. - Our own coverage. One site, a few page templates, one crawl on September 28, 2026 and two consecutive weeks of Search Console data compared side by side. ## **What does each failed audit check signal?** Checks 1, 2 and 6 usually need a developer or hosting administrator, and Checks 4, 5 and 7 need the agency or content team. Ask any agency which fixes it will do itself. Check A fail usually signals 1\. Bot rules A plugin default, a staging leftover or an unrecorded block. 2\. Schema A template gap, or markup that no longer matches the page. 3\. Entities Profiles that drifted apart after a rename or move. 4\. Chat assistants Thin branded coverage, or wrong facts assistants repeat. 5\. AI Overviews Ranking without citation, or no ranking. 6\. Crawlability A firewall rule stricter than robots.txt. 7\. Content format Answers buried under introductions. After you hire, hold the work to the seven monthly metrics in our monthly SEO report test. To have Mettevo run this AI visibility audit on your domain, [request an audit](https://mettevo.com/#contact). ## **FAQ** ### **How do you audit AI search visibility?** Run seven checks in order: robots.txt and llms.txt, schema.org markup, entity consistency, ChatGPT and Perplexity queries, AI Overviews for your top 20 queries, crawlability by AI user agents and content format. Google says AI Overviews need no special optimization, so the audit tests access and evidence. ### **What does an AI SEO audit include?** Seven checks: bot rules, schema.org coverage per template, entity profile consistency, ten manual queries in ChatGPT and Perplexity, AI Overview appearance for 20 queries, request tests for 10 agents that send requests and a content format review. An audit finds problems and evidence. It does not promise a citation. ### **How often should you run an AI visibility audit?** Run all seven once before you hire. After that, repeat the chat and AI Overview queries monthly, revisit entity profiles and content format quarterly, and rerun the bot rules, schema and crawlability tests after any release that touches robots.txt, templates or hosting. ### **Can you check brand citations in ChatGPT and Perplexity?** Yes, by hand. Write five branded and five category queries, run each three times in ChatGPT with web search on and in Perplexity, and log the date, whether your brand is named and which domains are cited. [OpenAI says](https://developers.openai.com/api/docs/bots) sites that opt out of its search crawler will not be shown in ChatGPT search answers. ### **What tools audit AI Overview citations?** Search Console's [Generative AI performance report](https://support.google.com/webmasters/answer/16984139) shows impressions in AI Overviews and AI Mode by page, country, device and date. Google says it reached all websites as of August 31, 2026. It counts impressions only, not clicks or queries. For cited domains per query, use a private browser window and a spreadsheet, or a tracker such as Ahrefs Brand Radar. ### **Is manual AI visibility checking reliable?** Partly. A manual run shows what appeared on that run and cannot prove absence. [Ahrefs measured](https://ahrefs.com/blog/ai-overview-change/) Google AI Overviews only and found a 70% chance of change between observations and 54.5% average overlap of cited sources (43,000 keywords, November 2025). Log each query three times and treat one miss as inconclusive.