Check whether an AI assistant can read your site, in an afternoon

You do not need a subscription to find out whether assistants can read your site. You need an afternoon and a spreadsheet.

The checks below are in order of how often they find something. The first one takes two minutes and fails more sites than the other five put together.

1. Turn JavaScript off

In Chrome: open DevTools, press Ctrl+Shift+P or Cmd+Shift+P, type “javascript”, choose Disable JavaScript, and reload.

If your content vanishes, you have found the problem and you can stop reading the rest of the list until it is fixed. A crawler that does not execute scripts gets exactly what you are now looking at. Plenty of modern sites assemble their content in the browser and serve an empty shell to everything else, and this is completely invisible in normal use.

This is not theoretical. The site this article is on was rebuilt partly because its predecessor assembled its project lists client-side. Half the portfolio did not exist unless you were a browser.

While you are there, check the page is not blank because of an animation. Text hidden at opacity zero waiting for a script to reveal it is present to a crawler but absent to a person, which is the opposite failure and also worth knowing about.

2. Read the page as text

Still with JavaScript off, select all and copy into a plain text editor.

Read what you get. Is the thing your business does stated in the first paragraph, in words a stranger would use? Or is the first paragraph a slogan, with the actual answer in the eighth?

An assistant lifts an answer. It cannot lift one that is not there, and it will not reward you for building to a reveal.

3. Confirm the bots are actually fetching

Check your logs for these user agents:

• GPTBot and OAI-SearchBot (OpenAI)
• ClaudeBot and Claude-User (Anthropic)
• PerplexityBot
• Google-Extended

On Cloudflare this is under the crawl controls, available on the free plan. On Vercel it is in the logs. On a traditional host it is the access log.

If none of them appear, nothing downstream matters. Check robots.txt has not blocked them, which is more common than you would think, often added years ago by somebody worried about scraping.

4. Turn on Bing Webmaster Tools

This is the single most useful free thing in the category and almost nobody uses it.

Bing Webmaster Tools has an AI Performance report, in public preview since February 2026. It shows total citations, which of your pages were cited, and, most valuable of all, grounding queries: the retrieval-side queries that led to you. It covers Copilot and Bing’s AI summaries rather than ChatGPT, and it is a sample rather than a census. It is still the only free place where you can see actual citations and the questions behind them.

Google Search Console added generative AI performance reporting in June 2026, but it gives impressions only: no clicks, no queries, no breakdown. Directional, not diagnostic.

5. Find your referral traffic

GA4’s default channel group now has an AI Assistant channel, covering arrivals from ChatGPT, Gemini, Copilot, Deepseek and Grok.

One trap worth knowing: Google’s own AI Overviews and AI Mode are not in that channel. They are counted as Organic Search. So the AI Assistant number is real but partial, and comparing it against “all AI traffic” will mislead you.

6. Build a prompt panel

This is the one that catches what the others cannot, because a citation that produces no click shows up nowhere else.

Write twenty to thirty questions a potential customer might actually type. Not your brand name. Questions like “who builds Webflow sites for fintech in Europe” or “how fast should a marketing site load”.

Then, on the same day each month, run every prompt in ChatGPT, Claude, Gemini and Perplexity, and log three things: were you cited, were you named without a link, or were you absent.

Two rules make this worth the hour it takes.

Keep the prompts fixed. The value is the comparison, so changing the questions destroys it.

Expect noise. These systems are non-deterministic; the same prompt can return different sources on two consecutive runs. One study ran 12 prompts across three tools with 600 volunteers, 2,961 runs in total, and concluded that measuring presence in AI answers with precision is a fool’s errand. The honest unit is a trend across months, never a position.

What to do with the results

If 1 or 3 failed, fix those and ignore everything else until they pass. They are bugs, and no amount of content works around them.

If 1 to 5 pass and the panel still shows you absent, the problem is almost certainly not on your site. Assistants corroborate: they draw on the places that talk about you more than the place where you talk about yourself. The fix is slow and it is mostly off your own domain, which is why so little of the advice in this category mentions it.

And if you are about to add an llms.txt because a checklist told you to: Google’s guide, updated 10 July 2026, lists it under things you can ignore, and a study of 137,210 domains found 97% of published llms.txt files were never requested at all. It takes five minutes, it does no harm, and it is not the thing standing between you and a citation.