The AI Visibility Audit: A 20-Point Checklist

What the audit is actually scoring

The stakes are a click-through problem before they are a brand problem. Pew Research Center tracked real Google users and found that when an AI summary appeared, people clicked a traditional search result on 8% of visits, versus 15% of visits without a summary — and only 1% clicked a link inside the summary itself. Being the source the summary is built from matters more than ranking below it.

So the audit scores five things in strict dependency order. Crawlability determines whether retrieval agents can read you at all. Entity consistency determines whether a model can resolve your name to one unambiguous organisation. Earned signals determine whether anything outside your own domain corroborates what you say. Citability determines whether your sentences survive extraction. Measurement determines whether you can tell any of this changed.

Score each point pass or fail with no partial credit. Half-credit is how audits turn into 40-page documents nobody actions. Twenty checks, twenty binary answers, then work the failures top-down. For the broader framing of why this discipline exists, see our primer on AI visibility.

Part 1 — Crawlability for AI (points 1–4)

Retrieval agents are not the same as training crawlers, and the distinction is where most robots.txt files go wrong. OpenAI documents three separate user agents — GPTBot for training, OAI-SearchBot for search indexing, and ChatGPT-User for live user-triggered fetches — in its bots documentation. Anthropic documents ClaudeBot and its block instructions in its own support article, and Perplexity publishes PerplexityBot and Perplexity-User. Blocking the training crawler is a legitimate business decision; blocking the retrieval agent by accident is self-harm.

Copilot answers lean on the Bing index, so if you are not verified in Bing Webmaster Tools you are flying blind on one of the six major assistants.

• 1. robots.txt permits retrieval agents. Check: fetch your /robots.txt and search for GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot, Perplexity-User, Google-Extended and Bingbot. Pass: every retrieval agent is allowed on commercially important paths, and any training-crawler block is a documented decision rather than a copied template.
• 2. No edge-level blocking. Check: curl your five most important URLs with each agent's user-agent string from a datacentre IP, not your office. Pass: HTTP 200 and full HTML — not a 403, a 429, or a bot-challenge interstitial from your WAF or CDN.
• 3. Content exists in the initial HTML. Check: view raw source, or fetch with JavaScript disabled. Pass: headline, body copy, pricing, specifications and key facts are all present before hydration. If your product page is an empty div, you are invisible regardless of robots.txt.
• 4. Sitemap, canonicals and speed are sane. Check: every revenue-relevant URL appears in the XML sitemap, returns 200, and carries a self-referencing canonical. Pass: no orphan pages, no canonical pointing at a redirect chain, and HTML delivered fast enough that a timeout-bounded fetch completes.

Part 2 — Entity consistency (points 5–8)

A model answering “who is X” is doing entity resolution against a lot of conflicting text. If you are “Acme”, “Acme Inc.”, “Acme Software” and “ACME” across four properties, and a different Acme exists in an adjacent market, the assistant will either hedge or blend the two. Blending is worse than absence: you inherit somebody else's pricing and somebody else's complaints.

Structured data is the cheapest disambiguation lever you have. Google's structured data documentation covers the mechanics, and the sameAs property is the one that does the identity work — it links your site to the external profiles that describe the same entity.

• 5. One canonical name string. Check: grep your homepage, footer, LinkedIn, Crunchbase, review profiles and press boilerplate for the legal and trading name. Pass: identical spelling, capitalisation and suffix everywhere, with variants used only as explicit aliases.
• 6. Organization schema with sameAs. Check: run your homepage through a structured data validator. Pass: a valid Organization (or LocalBusiness) node with name, url, logo, description and a sameAs array pointing at your verified external profiles.
• 7. Profile parity across platforms. Check: open every profile you control — LinkedIn, Crunchbase, G2, GitHub, app stores, local listings — side by side. Pass: same one-sentence descriptor, same founding year, same category, same address and phone, same URL format.
• 8. A resolvable public identity. Check: search your name on Wikidata and Wikipedia and see what comes back. Pass: either a correct Wikidata item with industry, founding date and official website populated per Wikidata's notability policy, or — at minimum — no rival entity outranking you for your own name.

Part 3 — Earned signals off-site (points 9–12)

Assistants build shortlists from third-party sources far more than from vendor sites. When someone asks for the best tool in a category, the model is usually synthesising listicles, review platforms, editorial coverage and community threads. Your own comparison page is a weak input in that mix; a roundup on a publication that already ranks for the query is a strong one.

This is the slowest category to move, which is exactly why it should be audited early. Discovering in month six that you are absent from every category roundup is a much more expensive discovery than finding a robots.txt error.

• 9. Category roundup coverage. Check: run the ten “best [category] for [use case]” queries your buyers actually use, list the top ten organic results, and record which of them include you. Pass: presence on more than half, with an accurate descriptor.
• 10. Review platform depth and recency. Check: on the two or three review sites that matter in your category, count total reviews, reviews in the last quarter, and how many contain substantive prose rather than a bare star rating. Pass: a steady recent flow and enough text for a model to quote a specific strength.
• 11. Independent editorial mentions. Check: search your brand name excluding your own domain and your press-release distribution. Pass: at least a handful of independent articles that describe what you do in a sentence a model could lift verbatim.
• 12. Community presence. Check: search your brand on Reddit, Stack Overflow, Hacker News and the niche forums your buyers use. Pass: real threads where customers describe outcomes — not just your own team posting, and not a wall of unanswered complaints.

Part 4 — Content citability (points 13–16)

Citability is a writing constraint, not a volume target. A model extracting an answer wants a self-contained proposition it can quote without the surrounding paragraph. Copy that warms up for three paragraphs before defining anything gives it nothing to take. Copy that opens with a flat declarative sentence containing the subject, the definition and a qualifier gets quoted.

The second constraint is specificity. “Fast setup” is unquotable. “Setup takes under 30 minutes for a standard Shopify install” is quotable, checkable and repeatable back to a user. Numbers, units, dates and named conditions all raise the odds that your sentence survives extraction intact.

The third is honesty about limits. Assistants are frequently asked comparative questions, and pages that state who a product is not for tend to be treated as more reliable than pages that claim universal fit.

• 13. Answer-first structure. Check: read the first 40 words of your ten most important pages in isolation. Pass: each one answers the question in the title as a standalone sentence, with no pronoun that needs the previous paragraph to resolve.
• 14. Facts carry units and dates. Check: highlight every claim on a key page and ask whether it could be verified by a stranger. Pass: quantities, timeframes, prices, limits and integrations are stated explicitly, with the date of any figure that changes.
• 15. Comparison and pricing pages state their own boundaries. Check: does any page say who the product suits badly? Pass: an explicit “not a good fit if…” section, and pricing that is stated on the site rather than gated behind a form.
• 16. Structure is machine-readable. Check: headings, tables, FAQ blocks and URL stability. Pass: H2s phrased as the questions buyers ask, specifications in tables rather than prose, stable URLs that do not change when the page is refreshed.

Part 5 — Measurement (points 17–20)

Assistant answers vary between runs even for identical prompts, so a single query is an anecdote. Measurement means a fixed prompt set, executed repeatedly across multiple assistants, with results stored so you can compare two dates and attribute movement to a change you made.

Build the prompt set from language buyers actually use — sales-call transcripts, support tickets, and the queries in your search console — not from keyword tools. Group them by intent stage: category discovery, shortlist comparison, objection handling, and branded queries. A shortlist prompt where you never appear is a different problem from a branded prompt where you appear with the wrong descriptor.

This is the part Growli automates: we run your prompt set across ChatGPT, Gemini, Claude, Perplexity, Copilot and Grok on a schedule, track mention rate, competitor share of voice, sentiment and the URLs each assistant cites, then rank the gaps by effort and impact. How it works covers the mechanics.

• 17. A documented prompt set exists. Check: is there a file with 30–100 prompts grouped by intent stage? Pass: prompts are in buyer language, cover unbranded and branded intent, and are version-controlled so changes are visible.
• 18. Multi-assistant, repeated runs. Check: how many assistants and how many runs per prompt per period? Pass: at least four assistants, several runs per prompt, and a stored history rather than screenshots in a chat thread.
• 19. Competitive and qualitative tracking. Check: do you record which competitors appear alongside you, and how you are described? Pass: named share of voice per prompt group, plus a flag for factually wrong descriptors — those are the highest-priority fixes in the whole audit.
• 20. Cited sources and referral traffic. Check: which URLs do assistants cite when they mention your category, and does your analytics segment referrals from assistant domains? Pass: a ranked list of the third-party pages that feed answers in your category, and a saved analytics view for AI-assistant referrers.

Scoring it and turning failures into a queue

Add up the passes. Sixteen or more and your problem is competitive, not structural — spend your effort on earned signals and on the specific prompts where a competitor dominates. Between 11 and 15 you have a mixed picture: fix the crawlability and entity items in week one, because they gate everything downstream. Ten or below and you should not be commissioning new content at all until the fetch path and the entity graph are clean.

Sequence matters more than the score. A crawlability failure invalidates every content investment made while it persists, and it is usually a one-line fix at the CDN. Entity failures take days. Citability rewrites take weeks. Earned signals take quarters. Working them in that order means each fix compounds on the last rather than being wasted on pages the assistants never fetch.

Re-run the full audit quarterly and the measurement points continuously. Crawler user agents change, CDN rules get tightened by someone in security, review volume decays, and category roundups get rewritten without you in them. The audit is a control, not a project.

FAQ