ChatGPT and Gemini APIs and interfaces overlap on sources by just 12–15%

SEO teams are used to a simple logic: allow a bot to crawl a page, wait for indexing, and then monitor rankings. With AI search, that model breaks down. Access to content, use of the material for model training, and the appearance of a link in a generated response are three different operations. A single switch does not control all of them at once.

A simple example: an online store allows Googlebot to crawl product pages, prohibits the use of its content for training, and wants to appear in AI Overviews. These decisions do not contradict one another, but they require different settings. If the team changes only robots.txt and sees no increase in citations a week later, that says nothing yet about the quality of the content or whether the restriction is working.

Access, training, and AI search must be separated

Google plans to support a common AIPREF-based preference signal for AI training in 2027. Until it launches, it cannot be treated as a universal setting already honored by every model. Existing rules from individual providers are not going away either. For example, Google-Extended controls some content-use scenarios, but it does not exclude a site from regular search and is not a ranking factor.

A separate mechanism has been announced for the generative features of Google Search. It is intended to let site owners manage their content’s participation in AI Overviews and AI Mode through Search Console. This setting is not the same as a training opt-out. Before changing it, teams should record baseline traffic, the set of pages involved, and the expected outcome of the experiment. Otherwise, they will get a chart but will not be able to explain what changed it.

Screaming Frog SEO Spider 25.0 simplifies the technical side of the process. Crawl configurations can now be stored as JSON, versioned, and reused. This is useful for recurring audits: the same set of URLs is checked with identical parameters, while configuration changes remain visible in the project history. However, a crawler confirms that a page is accessible, not the likelihood that it will appear in a model’s response.

Citations are becoming more measurable, but not automatically more accurate

Bing has added the Topics Portfolio tool to AI Performance. It groups queries by topic, displays citations, and suggests queries where a site may have a coverage gap. For editorial teams, this is more useful than an aggregate mention count: they can see that a brand is visible, for example, in the backup topic but is almost absent from the data security topic.

There is an important caveat. Clusters are generated automatically, and the values may be approximate. A small site may not see the tab at all because there is insufficient data. The opportunity list therefore should not be turned directly into an editorial plan. Teams must first review the queries manually, remove unrelated brands, and align the topics with the product and user intent.

Another limitation emerged from a study of AI product recommendations. In a sample of 1,536 responses, the overlap between domains shown in the user interface and through the API averaged 12.0% for ChatGPT and 14.8% for Gemini. This is not a universal benchmark for every industry: the study covered 117 English-language queries about physical products, run from the Netherlands. But the practical signal is strong: an automated API audit cannot be presented as an exact copy of the user experience.

What to change in the workflow

Start by creating an inventory. For each project, document crawl rules, training opt-outs, AI search participation settings, and available analytics sources. Then select a small, stable set of queries and test it simultaneously through the API and the actual interfaces. Record the date, region, query wording, model, links found, and visible citations.

Finally, include metadata in editorial acceptance checks. The title, description, JSON-LD, and alternative text should undergo the same fact-checking as the main article. The key conclusion is simple: a GEO audit becomes reproducible only when the team separates access settings from measurement methods. A single report, crawler, or API request does not provide the full picture.