How an audit works, step by step

You should know which numbers are measured and which are our own opinion. Here is the honest split.

We check that the address is a public website, then fetch:

• Your sitemap (sitemap.xml, or the one named in robots.txt). Its URLs are counted and classified; they are not all downloaded. For sitemap indexes we read up to 8 child sitemaps, preferring content ones (posts, pages, guides). We keep up to 5,000 URLs.
• Up to 10 pages: your homepage, common key pages (pricing, about, features, FAQ, docs, help, blog, case studies) and the most content-rich URLs from your sitemap. Reading stops after 50 seconds.
• Your robots.txt and llms.txt.

Pages are fetched directly. If that returns too little readable text (common on JavaScript-built sites or behind bot protection), we retry up to 5 pages through Jina Reader, which renders the page. Error pages served as normal pages (“404”, “Access denied”, bot challenges) are discarded, and pages that redirect to the same content are counted once.

If we can't read the site at all, the audit stops with a clear message before any AI call, so you aren't charged AI fees for an empty report.

Every number in a report is something we observed. Nothing is estimated, extrapolated, or added because a site is large.

The AI Readiness Score starts at 0. Each signal earns points depending on how close it is to a benchmark target, up to 100% of the target.

Signals well below their target earn less than proportional credit: at 60% of the target or more, full proportional credit; 30 to 60%, 70% of it; 10 to 30%, 40%; under 10%, 15%.

Each playbook action states the most it can add to one assistant's estimate if you reach its target (for example “up to about 9 points on the ChatGPT estimate”), worked out from the weights below. We don't show a combined “estimated improvement”, because not every gap applies to every business.

Benchmark targets are reference values set by us, adjusted for your industry (for example, restaurants and local businesses have much lower targets for tutorials and case studies, and law firms for tutorials) and scaled down for sites with fewer than 50 sitemap pages or no sitemap. They are not averages of real companies' data.

The report's industry benchmark card compares your scores with a reference and a stretch target for your industry (for example SaaS, legal, e-commerce, or restaurants and local businesses). These are also values we set, not measured averages.

Platform scores re-weight the same four categories:

These weights are our estimate of what each assistant tends to favour. They are not published by those companies, and platform scores are not rankings measured inside their products.

Grades: A+ 90+, A 85+, A- 80+, B+ 75+, B 70+, B- 65+, C+ 60+, C 55+, C- 50+, D+ 45+, D 40+, D- 35+, F below 35. Scores are never raised to a minimum, and the same site gives the same score whichever AI model you choose.

We parse robots.txt using the standard rules (RFC 9309): a crawler follows the group that names it, and only falls back to “User-agent: *” when none does. A crawler counts as blocked when its group disallows the whole site.

• AI search and answer crawlers: OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User, PerplexityBot, Perplexity-User, bingbot. Blocking these becomes the #1 action in your playbook.
• AI training crawlers: GPTBot, ClaudeBot, Google-Extended, Applebot-Extended, CCBot, meta-externalagent. Reported, but allowing them is your choice.

If robots.txt can't be read (some sites block our check), the report says “unknown” rather than assuming crawlers are allowed.

The AEO blueprint is a separate 0 to 100 checklist shown alongside the score: E-E-A-T signals, off-site authority, answer-first openings, schema completeness, freshness, AI bot access, heading coverage and entity consistency. It does not change the AI Readiness Score.

Off-site authority only uses checks we can verify:

• A Wikidata entity whose official website is your domain, and its Wikipedia article. If an entity with your exact name lists a different website, the report says so instead of giving credit.
• A LinkedIn company page, counted when your site links to it or the page exists at your brand's URL.

G2, Trustpilot, Capterra, Crunchbase and Reddit block automated requests, so we don't check them and never report them as missing. Check those yourself.

All AI calls run on your own OpenRouter key with the model you choose. AI is used for exactly these steps:

• Business details suggestion (1 call): reads the start of your homepage and suggests name, industry, audience and keywords. You confirm or edit them before the audit runs.
• Topic ideas and competitor comparison (1 call, retried once if the answer is unreadable): up to 20 questions in four categories, with more on request, up to 40 per category, plus a comparison if you named competitors. A topic that cites a number not found on your own site, or an outdated year, is rejected.
• AI Citation Check (optional, see below).

AI does not calculate scores, count signals, check crawlers, choose playbook actions or write the AEO Pack. Those are fixed rules, which is why results are repeatable.

1. You choose the engines: Perplexity, Gemini, Claude, GPT, Grok, or a web-search model of your own. You see the estimated cost first.
2. You choose the questions. By default your analysis model writes 8: 3 that name your business and 5 that do not. Or you write up to 25 of your own.
3. Every question goes to every engine you chose, through OpenRouter with live web search on. These are the companies' models through their APIs, so their consumer apps can answer differently.
4. We check each answer for your whole name or your domain. One more AI call lists the competitors named in each answer: businesses offering the same kind of product or service. Your own products, publications, marketplaces, delivery apps and tools that merely integrate with you are left out.

We count two groups apart. A question that already contains your name nearly always gets an answer naming you, so it says little about being recommended. The report leads with answers to questions that did not name you.

AI answers vary between runs and over time, so treat each check as a dated snapshot. The full answers are kept for your latest check, and a dated summary of each past check stays in your history.

The 10 files are generated from what we found on your pages: titles, meta descriptions, main-content paragraphs, a logo you declare in schema or as an app icon, social profiles you link to, and FAQs you already show. Nothing is invented; anything we couldn't verify is a clearly marked [REPLACE: …] placeholder.

llms.txt follows the llmstxt.org format. Schema follows schema.org; FAQ markup only uses questions visible on your site, in line with Google's structured-data guidelines. The robots.txt additions follow RFC 9309 and include a sitemap line only if we found your sitemap. llms.txt is a proposed convention: some AI tools read it, and its effect on citations is not yet proven.

• We read up to 10 pages plus your sitemap. Content that isn't in either can be missed, so a “0” means “not found in what we checked”.
• Platform scores and benchmarks are estimates and reference targets, not measurements from AI companies or other businesses.
• No search-volume, ranking or traffic data is used.
• Review sites (G2, Trustpilot, Capterra), Crunchbase and Reddit are not checked because they block automated requests.
• Sites behind logins or strong bot protection may not be readable.
• A better score does not guarantee AI assistants will cite you. It means you have fixed what we can measure.