The 30-minute AI visibility audit (checklist)

Open a fresh session in ChatGPT or Claude. Use a temporary chat or switch memory off, because past conversations about your own brand will contaminate the answers. Then ask five questions phrased the way a real buyer would phrase them, not the way you describe yourself internally. A useful mix: two category questions ('best accounting software for UK freelancers'), one alternatives question ('alternatives to [your biggest competitor]'), one problem question ('how do I automate invoice chasing') and one comparison ('[you] vs [competitor], which should I pick?'). Ask each in its own fresh session so one answer cannot influence the next.

Score each answer as you go: 2 points if you are named and recommended, 1 point if you are mentioned in passing, 0 if you are absent. That gives you a mention score out of 10.

Good looks like 6 or more out of 10, with at least one answer placing you in the top three recommendations. Below 4, assistants are effectively steering buyers elsewhere, and the causes are usually identifiable and fixable.

In another fresh session, ask two direct questions: 'What does [your brand] do and who is it for?' and 'How much does [your brand] cost?' Follow up with 'What are the main drawbacks of [your brand]?' Now check every claim against reality: current pricing, current product names, the audience you serve and any features it credits you with. Write down each error, however small.

Expect to find some. When we tested 30 brands in July 2026, assistants misdescribed 27 of them (90 percent) when answering from memory, and errors survived in 13 of 30 answers (43 percent) even with live web search switched on. The full breakdown is in our study of how often AI gets brands wrong.

Good looks like no material errors: the pricing is right, and nothing retired or invented is presented as current fact. Stale but harmless details, like an old tagline or a slightly dated customer count, matter less than claims that would mislead a buyer.

Open yourdomain.com/robots.txt in a browser. Look for Disallow rules under AI user agents such as GPTBot, ClaudeBot, PerplexityBot and Google-Extended. A 'Disallow: /' under any of these means that crawler cannot read your site at all. Then try yourdomain.com/llms.txt, a plain-text summary file some sites now publish for AI systems. If you are not sure whether you need one, start with what llms.txt is.

If you would rather not read robots.txt syntax line by line, the free AI crawler checker does this step in one pass: it tests your robots.txt against 12 AI crawlers and checks whether an llms.txt file exists.

Good looks like deliberate access. Blocking AI crawlers is a legitimate choice for some businesses, but it should be a decision someone made on purpose, not a leftover from a CDN default or an old bot-blocking plugin. If you find a block you cannot explain, that is your first fix.

Search Google for 'best [your category]' and one qualified variant, such as 'best [category] for small businesses UK'. Open the top five organic results for each search and count how many name you. These round-up pages matter because assistants with live search read and cite them heavily; if the leading listicles in your category omit you, live-search answers tend to omit you too. Good looks like appearing on at least three of the top five, with your pricing and positioning described correctly.

Next, check the review sites that count in your category. For software that usually means G2 and Capterra; for consumer services, Trustpilot. Add whatever directory your industry actually uses. For each profile, confirm four things: it is claimed, the description matches your current website, the pricing is right and the newest review is under six months old.

Good looks like every relevant profile claimed and current. A profile still describing your 2023 product line does active damage, because assistants treat these pages as authoritative sources about you.

Put the results in a spreadsheet with one row per audit: date, mention score out of 10, number of factual errors found, crawler check pass or fail, best-of page count out of five, and whether your review profiles were current. Keep the exact wording of your five questions from step 1; changed questions make next month's numbers incomparable.

A single audit tells you where you stand today, but the comparison is where the value is. Assistant answers vary between runs, so treat any one result loosely and watch the direction across months instead. If you fix something in the meantime (unblock a crawler, say, or update a stale G2 profile), the next audit shows whether the answers actually moved.

If the manual loop is more discipline than you want to sign up for, the free AI visibility checker is the automated version of steps 1 and 2: it asks buyer-style questions about your brand and fact-checks the answers, running through Claude with live web search (other engines are rolling out), with no login required. Paid plans re-run the same check weekly and keep a score trend history, which replaces the spreadsheet. Either way, record a baseline now; it is the only way to know whether anything you change later shows up in the answers.