Most teams start an AI search audit by typing their company name into ChatGPT. It is the cheapest test and usually the least useful one. A model that knows your brand can still recommend a competitor when a buyer describes a problem instead of naming a vendor.
A useful audit starts earlier, with data you already own: search queries, Bing citations, visits from AI assistants and sales conversations. That data gives you the questions worth checking in the models, and the result becomes a plan for whoever edits your pages. Here are the five steps, where the data for each one lives, and what it costs.
Step 1: Find buyer questions in Google Search Console
Search Console will not show you what people type into ChatGPT. It will show how they search for your category on Google, and those questions often reappear in AI assistants almost word for word.
Open Performance → Search results, set the last 3–6 months and add a query filter with a regular expression. Search Console uses RE2 syntax (Google help). Two filters to start with:
^(how|what|which|why|is|can|does|best|vs) ^(\S+\s){5,}\S+$The first catches queries phrased as questions. The second shows queries of six words or more, where buyers describe their situation: "invoicing software for accounting firms with QuickBooks sync". That is the closest thing to a prompt you will find in search data.
Note the ranking page for each query. Queries with many impressions and a low CTR are good candidates: the topic matters, but your answer is not winning.
One limitation: Google counts AI Overviews and AI Mode in the overall Web data and does not split them out (Google documentation). A falling CTR with stable impressions may be caused by an AI answer above the results, but Search Console cannot confirm it.
Step 2: Check citations in Bing Webmaster Tools
Bing Webmaster Tools has the only free report that shows your pages being cited in AI answers. AI Performance entered public preview in February 2026 and covers Microsoft Copilot, AI summaries in Bing and select partner integrations (Bing announcement).
It shows:
- total citations and average cited pages per day,
- page-level citation counts,
- grounding queries: short queries the model writes itself to retrieve sources for an answer.
Grounding queries are the valuable part. They show how the model rephrased the user's request before searching. Compare them with your Step 1 list: queries that appear in both places but lead to someone else's pages go to the top. Since June 2026 Bing also classifies intents and shows citation share per query (Bing).
Know what the report does not tell you. According to Bing's help page, a citation is not a click or a visit, and the data is a sample, not a complete log. It does not cover ChatGPT, Gemini or Claude either.
Step 3: See who already arrives from AI: GA4, logs and CRM
This step answers whether AI already sends you visitors and which pages benefit.
GA4. In the acquisition reports, filter session source for chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com or claude.ai. Some ChatGPT visits carry utm_source=chatgpt.com without a utm_medium, so GA4 may file them under "Unassigned" rather than "Referral". The reliable fix is a custom "AI" channel group placed above the Referral rule and matched on source only. Record the landing pages: they are the pages models already cite.
Server logs. Check whether AI crawlers fetch your pages at all. For ChatGPT, three agents matter (OpenAI documentation):
OAI-SearchBotcontrols presence in ChatGPT search. Ifrobots.txtblocks it, your pages will not appear in search answers (apart from navigational links).ChatGPT-Uservisits a page when a user asks for it in a conversation.GPTBotcollects training data. You can block it without affecting search.
A robots.txt change takes about 24 hours to apply in ChatGPT search. If you cannot access logs, ask your host, or check what Cloudflare or your CDN records.
CRM and sales calls. Add a "how did you hear about us" field to your forms and read recent demo notes. "ChatGPT recommended you and two others" tells you more than a chart: who was on the list and for which question.
Output of Steps 1–3: a sheet with 10–20 buyer questions. For each: the exact wording, the page that should answer it, the data source (GSC, Bing, CRM) and whether AI already cites you or a competitor there.
Step 4: Run the questions through the models and compare sources with your page in TextRank
You have the list. Now you need to know what ChatGPT, Gemini, Claude and Grok answer, which sources they rely on and what your page is missing.
By hand, that means asking each question in each model, copying the answer and links, opening every source next to your page and writing down the differences. With 10 questions and two models, that is 20 answers and usually dozens of sources to read.
In TextRank you enter the question and your page URL, and the audit:
- collects the selected model's answer and the sources it cites or mentions,
- checks whether the model can fetch your page,
- compares the sources with your page at the level of entities, claims and standalone passages,
- assigns each gap a move (keep, rewrite, add an answer, get proof, create supporting material or park the prompt) with a priority and a reason.
If you do not have the list from Steps 1–3 yet, the prompt module suggests buyer questions based on your site. For choosing them yourself, see how to choose buyer prompts for an AI audit.
Step 5: Hand the plan to the person or AI that edits the site
An audit is worth the changes it gets onto the page. That is why TextRank outputs a plan rather than a chart: each recommendation points to the prompt, the model's answer, the source and the exact passage on your page. You export it to Markdown or PDF.
Two ways to implement it:
- A person with CMS access (copywriter, product marketer, site owner) gets the PDF and works through the priorities. They do not need to understand AI search. They need to know which passage to change and why.
- An AI tool (Claude, ChatGPT, an agent in your site repository) gets the Markdown file as context along with the page content. Ask for a draft change for one recommendation at a time, and do not let it add facts that are not in your documentation.
Either way, product facts are approved by someone who knows the offer. The audit can show that a competitor states its implementation time; it cannot tell you yours.
After publishing, run the same audit again and compare sources with the first result. A single new mention proves little, because model answers vary between runs. Look for a lasting change across checks.
What an AI search audit built this way costs
| Item | Tool cost | Team time (assumption) |
|---|---|---|
| Google Search Console | $0 | 1–2 hours for filters and the list |
| Bing Webmaster Tools (AI Performance) | $0 | 1 hour |
| GA4, logs, CRM | $0 (log access depends on hosting) | 1–3 hours |
| TextRank audit: 10 questions × ChatGPT | 10 audits | a few minutes to run |
| Retest after changes | 10 audits | same |
| Implementing changes | $0 or copywriter cost | about 1 hour per page plus approval |
The hours are assumptions for your own estimate: use the rate of whoever does the work.
A $25 starter pack with no subscription covers roughly 4–7 audits, depending on the model. Premium at $99 per month covers roughly 20–35 audits a month. See current pricing.
For the example above (20 audits including the retest) you have two options: a few starter packs with no subscription, or one month of Premium with room left to check the same questions in another model.
For a one-off audit, packs are cheaper. If you return to the same questions every month and add a model, Premium pays off. For a comparison with agency quotes and monitoring tools, see AI brand visibility audit pricing.
What this process does not cover
If every cited source is a review site, a ranking or a forum, edits to your own pages will help less than being present in those places. The audit will show that pattern, but it will not earn reviews for you. It does not replace technical SEO either: Google notes that the usual foundations apply to its AI features. And no one can honestly promise that a model will start recommending you by a given date.
Where to start today
Open Search Console, add the long-query filter and pick one question where your page gets impressions but a weak CTR. Check it in TextRank and hand the first recommendation to whoever edits that page. To see a full result, read the example audit from prompt to page change.