Leadership asks for an “AI visibility analysis.” A week later they get a deck: 30 ChatGPT screenshots, a brand share chart and the conclusion “competitors get recommended more often, so we need more content.” Nobody knows which content, for which questions, or why. Two months later nobody can check whether anything changed either, because the conditions of the first measurement were never recorded.
TextRank provides competitor content and available AI-source analysis to support a visibility audit. Its deliverable is a specific plan: what to change on your page, where, why and in which order. A broader audit may also include monitoring, technical review and traffic and conversion data.
You can recognize a professional AI source analysis by one trait: every recommendation can be traced from the buyer’s question, through the model’s answer and the cited source, to a specific passage on your page. Everything else follows from that: the methodology, the levels of analysis, the way findings are recorded and the retest.
Why methodology isn’t a formality
The best example comes from the industry itself. In July 2025, Ahrefs reported that 76% of pages cited in Google AI Overviews ranked in the top 10. In March 2026, the same company reported about 38%. The authors attribute part of the gap to a change in how they detect citations, and part possibly to the model change, since Gemini 3 has powered AI Overviews from January 2026. Ahrefs is upfront that the figures aren’t directly comparable.
If the picture can shift that much in a large study, then in a single company’s analysis, differences in date, question wording or model mode can reverse the conclusions entirely. SparkToro and Gumshoe.ai found that ChatGPT and Google’s AI returned the same brand list less than once in 100 runs. Rand Fishkin also called tools that report an “AI ranking position” misleading.
That leads to three methodological requirements:
- Questions chosen for a reason. Ideally drawn from sales calls, support tickets and customer questions, each tied to a buying decision and to the page that should answer it. See how to choose buyer prompts for an AI audit.
- Recorded conditions. Question wording, language, model, mode, date and number of runs.
- Models reported separately. An average of ChatGPT and Gemini describes neither.
Five levels of analysis
An amateur analysis stops at domains: “they cite G2, Capterra and competitor X.” A professional one goes further down:
- Domain and source type. Your site, a competitor, media, reviews, documentation, forums, video. This tells you where the category conversation happens.
- URL. Which page exactly: pricing, a comparison, an article, a thread.
- Entities. Which companies, products, features, standards and external systems the source names, and how it describes the relationships between them.
- Claims. Which specific statements the model carries into its answer: “setup takes two weeks,” “the integration syncs both ways.”
- Passages. Which paragraph, table row or FAQ answer holds that claim, and whether it works without the rest of the page.
Editorial conclusions come from levels 4 and 5. Levels 1–3 tell you where to look.
A named gap instead of generic advice
“Add more content about integrations” isn’t a finding. A finding is a named gap with its reasoning. A professional report distinguishes between types of gap:
| Gap type | Example |
|---|---|
| Missing question | The page doesn’t say whether the product supports SSO through Okta |
| Missing fact | The page promises “fast setup”; sources give a number of days |
| Missing evidence | A GDPR compliance claim with no link to a document |
| Unclear entity | The page says “our integration” without naming the system |
| Unusable format | Prices only in an image, terms only in a PDF |
| Contradiction | The page and the docs state different limits |
Each gap gets a decision: keep the passage, rewrite it, add evidence, create a separate piece or park the question. Parking is a decision too. Some questions aren’t worth the work, because they are dominated by sources the company can’t influence, or because they sit far from a purchase.
What one entry in the report might look like
Say the company sells fleet management software to trucking firms. One entry in a professional report might read:
- Question: “Fleet management software with tachograph integration for a company with 40 trucks” - Models and date: ChatGPT with search and Gemini, 3 runs each, October 14 - Observation: 5 of 6 answers cited competitor A’s page, each time for a passage on automatic digital tachograph downloads and the number of supported device models. - Your page: The “Integrations” section lists “tachographs” with no device models, download method or frequency. - Gap: missing fact and unclear entity. - Decision: rewrite the section and add the list of supported devices and download frequency. The product team confirms the facts. - Priority: high: the question is close to a decision, the change is on an existing page, the facts are available internally. - Retest: same questions and models four weeks after publishing.
An entry like this can go to the writer, the expert and the approver without another meeting.
Amateur vs professional analysis
| Element | Amateur analysis | Professional analysis |
|---|---|---|
| Questions | Invented by the team, generic | From customer conversations, tied to a decision and a page |
| Measurement | One run, a screenshot | Repeat runs, recorded conditions, models kept separate |
| Unit of analysis | Domain | Claim and passage |
| Output | A share chart or a single score | A named gap with reasoning and a decision |
| Recommendation | “Create more content” | “Rewrite section X, add fact Y confirmed by Z” |
| Verification | None | Retest under similar conditions |
| Metrics | Everything in one number | Citations, mentions, visits and sales inquiries kept separate |
The most common problems with AI source analyses
A single score with no reasoning. A 62/100 doesn’t tell you what to fix. It can help set the order of work, but it can’t replace a description of the gap.
Ranking positions in AI answers. With answers this variable, list position is a weak measure. How often you appear across many answers is more stable.
Recommendations with no source for the fact. The analysis suggests adding “7-day setup” although nobody at the company has confirmed it. An unsupported fact must be flagged and verified before publishing.
Applying research findings to one company. The Princeton GEO study showed that the effectiveness of optimization methods varies across domains. General data tells you where to look, not what will work for you.
Not separating citations from business results. A citation in an answer, a visit to your site and a sales inquiry are three different events. Google reports traffic from AI features in Search Console together with other web search traffic, so attribution needs separate attention.
How long a professional analysis takes without tools
Take 15 questions, four models and three runs each. In our working estimate, collecting answers with proper records alone takes about 15–20 hours. Reading sources and filling in passage cards takes another 25–40 hours. Comparing against the client’s pages, naming gaps and setting priorities takes 10–15 hours, and writing the report and walking the team through it another 5–10.
That totals 55–85 hours: roughly two weeks of an analyst’s full-time work, or three to four weeks alongside other duties. A retest after changes go live repeats most of that work. That’s the main reason companies run this kind of analysis once and never come back to it.
How TextRank delivers this standard
TextRank was built so this standard doesn’t require weeks of manual work. For a chosen buyer question, it collects answers and cited sources from ChatGPT, Gemini, Claude and Grok, separately for each model, with the prompt and date recorded. It analyzes your page at the level of entities, claims and passages and compares it with the sources.
The report includes what we described above as the professional minimum: model answers, cited sources, a passage-by-passage analysis of your page, a named gap and a prioritized change plan. Every recommendation carries the prompt, the reasoning and the next step. You can share the report or export it to Markdown or PDF, then re-run the audit after the changes and compare.
TextRank doesn’t replace your team’s judgment. It doesn’t confirm product facts for you, and it doesn’t promise that a model will start recommending your company. It leaves your team the part not worth automating: deciding what is true and what to publish. The method is described in detail in our documentation, and you can see the full path from question to change in this example AI visibility audit. If you’re starting with the citations themselves, read how to analyze the content ChatGPT cites.
Run your first source analysis in TextRank.
Turn the analysis into a task for your writer
After comparing sources, name the target URL, passage, missing information, reasoning and acceptance condition. See how to hand over a competitor-content edit plan. The writer prepares the passage, a product expert approves facts and the page owner publishes the change.