ChatGPT Comparison Citations Rose 31.9 Points to 88.5%

By Nathan Brooks

Logged-out ChatGPT responses with citations rose from 56.6% in the week of Aug. 3 to 88.5% in the week of Sept. 28, according to Graphite’s tracked comparison prompts.

The 31.9-point increase sat alongside Graphite’s central Oct. 16 finding: low visibility without search generally accompanied low visibility with search. This held even when those brands were substantially mentioned on at least 15% of cited pages.

Among brands substantially mentioned on at least 15% of cited pages, median visibility with search was 11% when ChatGPT named them in fewer than 25% of its no-search answers. Graphite defines visibility as the share of responses mentioning a brand or product.

Graphite argues that brands should build training knowledge as well as citations. It calls the approach “parametric knowledge optimization (PKO).”

What did Graphite compare?

The analysis drew on entity-comparison prompts from Graphite’s Aug. 20 paper. It covered 3,115 entity-question pairs across 57 questions. Graphite collected the data in August from logged-out ChatGPT.

Each prompt asked ChatGPT to compare brands or products. For search answers, Graphite measured the share of cited pages substantially mentioning each brand, then split brands into three groups. In every group, brands named more often without search also appeared more often with search. Search visibility rose alongside the cited-page mention share.

The Oct. 16 write-up does not explain how Graphite isolated responses that did not use search. The Aug. 20 paper compared forcing ChatGPT to use search with not forcing it. It classified search use when a response contained a fan-out query or search result. Only the separate 100-prompt API replication explicitly disabled search.

Graphite reported the same relationship, with comparable strength, in two replications. One used 100 API prompts with search explicitly enabled and disabled. The other tracked 200 prompts through Graphite’s tracking tool during a different three-week window.

The study offers only early evidence that training knowledge affects how models use search results. Graphite did not describe controls for baseline brand awareness. Visibility with and without search moved together.

How did ChatGPT add brands to background searches?

Graphite counted 957,000 ChatGPT fan-out queries, the background searches run before an answer. Graphite collected the queries from 22,953 non-branded entity-comparison prompts between August and early October.

None of the original prompts named a brand. Yet 59.6% of fan-out queries contained one, according to labels from Anthropic’s Claude Haiku 4.5 model. Another 22.7% used a site: operator.

Graphite says a brand in a fan-out query raises the chance of retrieving pages featuring that brand. That is a retrieval hypothesis, not a controlled causal estimate.

In an August SEJ test, Suganthan Mohanadasan found a similar pattern. In 21 of 27 conversations on one logged-in account, ChatGPT’s first search query included brand names he had not typed.

What do the citation rates show?

For logged-out Gemini, Graphite recorded 79.0% citations in the week of Aug. 3 and 81.2% in the week of Sept. 28. ChatGPT’s 31.9-point rise was much larger.

Graphite interprets ChatGPT’s increase as evidence of more frequent search use. It also says these rates can fall again. A citation does not prove a search occurred. Graphite noted in May that ChatGPT could add inline links without showing evidence of a search.

Ahrefs reported on Oct. 12 that 14.4% of 66.6 million free, logged-out ChatGPT responses in August and September included fan-out queries. The figures do not align directly. Ahrefs used general prompts selected through search volume across all topics and countries. Graphite measured citations on fixed comparison prompts.

What should brands optimize?

Graphite is a growth agency that sells AEO services. Its PKO strategy treats training knowledge and citations as connected levers.

Graphite calculated days between each model’s publication date and vendor-reported knowledge cutoff. For Anthropic, it used the training-data cutoff rather than an earlier reliable knowledge cutoff, allowing comparison across three AI vendors.

For models released since July, those intervals ran about two to five months. Graphite calls this “the minimum time-to-impact for newly published content.” The calculation does not measure when a page enters the training set or whether model answers change.

Graphite’s proposed methods remain exploratory. One checks brand visibility through the API after each model launch, with search off and at least 50 responses collected per prompt. Another permits training crawler access and checks whether CDNs or firewalls restrict them.

Google’s Google-Extended robots.txt token offers a concrete control for Gemini. Google says it governs whether crawled content can train future Gemini models and support grounding in Gemini Apps and Grounding with Google Search on Vertex AI. It does not affect Google Search inclusion and is not a ranking signal there.

What does this mean for brand sites?

Citation share and cited-page mention share omit a key outcome: how often ChatGPT names a company without searching. In Graphite’s comparison prompts, low no-search visibility remained low during searches, even when brands appeared on at least 15% of cited pages.

For brands, strong citation coverage alone may overstate expected visibility when baseline model awareness is weak. Graphite recommends API tracking with search off as a baseline after each model release. Sites behind restrictive CDNs or firewalls also need to check training crawler permissions.

The data show correlation, not causation. They do not prove that increasing cited-page mentions or no-search visibility will increase brand mentions when search runs. Search-on tracking still needs repetition as behavior changes.

Two signals merit monitoring next: model releases and citation rates. Graphite says to expect a two-to-five-month gap between cutoff and release for new material to appear in an AI model. Because labs do not disclose their training data, brands cannot know which pages were included.