AI visibility audit

What can AI assistants actually read on your site?

Readback is an AI visibility audit. It requests your page the way ChatGPT, Claude, Perplexity and Google do, runs 56 checks on what comes back, and hands you a ranked list of fixes with the code to paste.

Scored on evidence, not folklore

Each pillar’s weight follows published research on what AI systems cite, and the crawler documentation from OpenAI, Anthropic, Perplexity, Google and Microsoft. Being reachable carries 26% of the score. llms.txt carries under 1%, because nobody has yet shown that it changes anything.

26%

Crawler access

Can AI crawlers and live fetchers reach the page?

URL accessibility scores 9.5 and preview controls 9.2 out of 10 in the Zyppy review of 54 studies. Nothing else matters if this fails.

18%

Answer structure

Can a model lift a clear, self-contained answer?

Query-answer match 9.2, answer near the top 8.8, AI-ready structure 8.6, self-contained passages 8.0.

14%

Content visibility

Is the content in the HTML that crawlers receive?

Content visibility 7.6. OpenAI, Anthropic, Perplexity and Meta crawlers do not run JavaScript.

10%

Evidence and specificity

Are there concrete facts and sources worth citing?

Factually specific 8.3, explicit phrasing 8.1, cites sources 8.0.

10%

Entity clarity

Is it unambiguous who is behind the site?

Brand and entity trust 6.8, entity consistency 5.8. Brand mentions are the strongest correlate in Ahrefs data.

8%

Freshness and trust

Does the page show when it was updated and by whom?

Freshness 7.0. AI-cited pages are about 26% fresher than organic results in Ahrefs data.

8%

Structured data

Is the page described in machine-readable form?

Structured data 5.6: a small but consistent positive relationship.

4%

Technical hygiene

Are URLs, redirects and status codes clean?

Supports crawling and indexing, which the top-ranked factors depend on.

2%

AI-native extras

llms.txt, Markdown responses, content signals.

llms.txt scores 2.0, the lowest of 23 factors. Google says no AI-specific files are needed. Weighted accordingly.

What a report gives you

One page is audited in depth and up to five more are sampled. Everything below is produced from your site’s own responses, in about twenty requests.

A ranked action plan

Every failed check ordered by the points it is worth, with the reason it matters, what was found on your page and the steps to fix it.

Crawler access, tested twice

22 AI crawlers checked against your robots.txt. Eight of them are then sent as live requests, which catches firewall blocks that robots.txt hides.

What AI sees

The text a non-rendering crawler extracts from your HTML, so content that only exists after JavaScript runs is obvious at a glance.

A verdict per assistant

Open, limited or blocked for ChatGPT, Claude, Perplexity, Google AI Overviews and AI Mode, Gemini, Copilot and Meta AI, with the reason.

A fix kit

A robots.txt block, Organization and Article schema, missing head tags and an llms.txt draft, generated from your own pages and ready to paste.

Exports

Copy the whole report as Markdown for a ticket or a client, or save the raw scan as JSON.

How scoring works

The score estimates how ready a page is to be found, read and quoted by AI systems. It is built from what your server returns, so the same page gets the same score every time.

  1. CollectThe page is fetched as a browser, then as eight AI agents. robots.txt, the sitemap, llms.txt and a deliberately missing URL are requested as well.
  2. Check56 checks run on the raw HTML and response headers, with no JavaScript, which is how most AI crawlers read a page. Checks that do not apply to the page type are left out.
  3. WeighEach pillar scores from 0 to 100 and is weighted as shown above. Blocking problems cap the total: a page that assistants cannot fetch cannot score well on writing quality.

What this audit cannot tell you

Readiness is not results

Whether assistants cite you also depends on how you rank for the searches they run and how often other sites mention your brand. Those are the strongest factors in the research, and neither is visible from your own pages.

Live tests are imitations

Agent requests are sent from this tool, not from the AI companies’ networks. A firewall that verifies crawler IP addresses can treat the two differently, so confirm blocks in your server logs.

Measured data lives elsewhere

For real citation counts use the AI Performance report in Bing Webmaster Tools and the Performance report in Google Search Console, and spot-check answers in the assistants themselves.

Sources behind the weights

Audit pasted page source

Useful for pages that are not public yet. Open the page, choose View Source, and paste the HTML. Network checks (status codes, firewall tests, speed) are skipped.