You ask ChatGPT, “Which invoicing software should a software agency with EU clients choose?” The answer links four sources: a competitor’s comparison page, an article on an accounting site, a forum thread and a YouTube video. The team’s first reaction is “we need a page like the competitor’s.” Usually that is the wrong conclusion, because nobody knows yet what that page was cited for.
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.
A cited source isn’t a template to copy. It is evidence that someone answered a specific sub-question better than everyone else. The analysis has to find which sub-question, which passage and which fact. Only then do you know what to add to your own page.
Start by logging the conditions
Model answers are unstable. Rand Fishkin of SparkToro and Patrick O’Donnell of Gumshoe.ai collected nearly 3,000 responses from ChatGPT, Claude and Google’s AI to 12 brand recommendation prompts. In ChatGPT and Google’s AI, the chance of getting the same list twice was less than 1 in 100, and the same list in the same order less than 1 in 1,000. The authors say they are not professional researchers and invite others to repeat the work. The takeaway was still clear: how often a brand appears across many runs of similar prompts is more stable than its position. Search Engine Journal has a summary.
So every observation needs a record:
- the exact question and its language;
- the model, version or mode (for example, ChatGPT with search);
- date and time, location, and whether you were signed in;
- the full answer and every link, including footnotes;
- the run number.
Without this, you can’t compare results after a page change with results before it. You’re comparing two dice rolls.
Separate mentions, recommendations and citations
These are three different events. A mention means your company’s name appears in the answer. A recommendation means the model points to your company as an answer to the question. A citation is a link to a page the model used for part of the answer.
Your company can be recommended based on an industry article, with no link to your site. Your page can be cited as the source of a definition while a competitor gets recommended. Content analysis focuses mainly on citations, because only they show which text the model actually used.
Classify the sources
Before reading anything, assign each source to a category: your own site, a competitor’s site, media and industry publications, review sites and directories, documentation, user-generated content (forums, Reddit) and video.
The mix tells you something on its own. If review sites dominate a vendor-selection question, fixing your product page may not be enough. If competitors’ documentation dominates, your docs probably lack specifics. Don’t ignore video: in Ahrefs’ March 2026 analysis, YouTube was the most cited domain in Google AI Overviews and accounted for 18.2% of cited URLs that ranked outside the top 100.
Break the cited page into passages
This is the most important and most time-consuming stage. You aren’t rating the whole page. You’re looking for the passage that matches the model’s answer, usually by a telling number, name or phrase.
For each passage you find, record:
| Field | Example |
|---|---|
| Sub-question it answers | Does the software handle EU reverse-charge VAT? |
| Passage text | “Reverse-charge invoices are generated automatically after the VAT number is validated in VIES” |
| Type of information | Feature + condition + named external system |
| Evidence | Documentation link, screenshot |
| Does it stand on its own | Yes: it names the product and the feature |
| What your page says | “Supports international invoicing,” no details |
A card like this turns an impression (“the competitor has a better page”) into a named gap (“we don’t explain how we handle EU VAT”). Google says its AI features may run related searches across subtopics to build an answer. That is why a single model answer is often stitched together from several passages on different pages. Map each one to its own sub-question.
Compare models separately
ChatGPT, Gemini, Claude and Grok use different retrieval systems and may cite different sources for the same question. Don’t average them. A source that matters for Gemini may mean nothing for ChatGPT. Keep results separate and look for patterns: if the same fact shows up in sources for three models, it probably matters for that question.
The most common analysis mistakes
Drawing conclusions from one run. One answer is an anecdote. Work from several runs and from frequency.
Confusing mentions with citations. “We appear in 40% of answers” doesn’t tell you whether the model used your page.
Analyzing domains instead of passages. “They cite G2” isn’t an editorial finding. “They cite a review where customers describe setup time” is.
Copying form instead of substance. The competitor has a table, so the team adds a table. But the model took one number from it, and your table still doesn’t give that number.
Not checking your own page is reachable. If robots.txt blocks OAI-SearchBot, your page won’t appear in ChatGPT search answers however good the content is.
How long it takes by hand
Take one question and four models. Collecting answers with a proper record takes about 20–30 minutes. With a few cited pages per model, some overlapping, you are left with a dozen or so unique sources to read. Finding the right passages and filling in the cards takes one to two hours. In our working estimate, one question is about 1.5–3 hours of careful work.
Twenty questions is 30–60 hours per cycle, without repeat runs. With three runs for reliability, answer collection triples. After you ship changes, the whole cycle has to run again to see the effect. In practice, teams without tools check a handful of questions once and make decisions on that.
How TextRank runs this analysis
TextRank automates the stages that take the most time by hand. For a chosen buyer question, it collects answers and cited sources from ChatGPT, Gemini, Claude and Grok, separately for each model. It logs the prompt, model and date, so the next audit compares results under similar conditions.
It then compares the sources with your page at the level of entities, claims and passages. Instead of a list of domains, you get a named gap (a missing fact, piece of evidence, condition or format) and a recommendation: keep, rewrite, add evidence, create a separate piece or park the question. Every recommendation traces back to the source that justifies it.
The tool doesn’t decide whether a fact is true for your company. That stays with your team. To see what a full report from this kind of analysis should contain, read what a professional AI source analysis looks like. For choosing questions, see how to choose buyer prompts for an AI audit, and for writing the passages themselves, how to write content for AI SEO and GEO.
Analyze the sources for your question in TextRank.
Tools for owned-page and third-party source review
Choose according to whether you need to compare your page, inspect a cited URL, or investigate a third-party publication. These are separate assignments. Discovering one domain does not automatically justify a new article or campaign: establish the information gap and appropriate action first.
- Top 5 GEO content analysis tools for sources, passages, and edits
- Top 5 tools for analyzing your website's AI citations
- Top 5 tools for discovering external sources cited by AI
- How to compare the same page in three SEO and GEO tools
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.