ChatGPT Citation Experiments (2026): 45 Figures From StatWharf Test Runs

From September 8 to October 11, 2026, StatWharf sent software category prompts to AI answer engines. This page reports what those runs measured. It covers four topics. The first is how often each engine cited a StatWharf page. The second is the searches ChatGPT ran before it answered. The others are prompt wording and how stable a citation is.

Every figure comes from StatWharf’s own experiment records. No third-party citation dataset is used. Each figure states the run date, the engine and the sample size (n). Most samples are small. The figures describe what happened in these runs. They do not show what causes a citation, and they are not stable citation rates for all users.

StatWharf publishes a longer write-up of one of these runs at How ChatGPT searches for software recommendations.

Method

StatWharf ran 9 experiment runs. Each run sent fixed prompts to an answer engine through a browser. It saved the full answer and the cited links. Where the engine exposed its searches, the run saved them too. Every prompt opened a fresh conversation. No prompt asked the engine to search, to cite or to mention StatWharf.

The core prompt panel used 50 software categories that StatWharf covers. Each category had a baseline prompt, best {category}, and a year prompt, best {category} 2026. Later runs used subsets of these categories and added new wordings.

A run counts as a citation only with a statwharf.com link among the sources the engine showed. The “intended page” is the StatWharf page for the category in the prompt. Failed captures, bot walls and answers saved before they finished are excluded from every count. The records, scripts and per-run tables are kept in the StatWharf experiment archive, dated by run.

Limits

Runs used one location in India and one account per engine, so results reflect that setup. Several answers localized to India without being asked. Most ChatGPT runs used model routing set to auto. The September 8 run was logged out, so it is not a clean comparison with later signed-in runs. Prompts are synthetic. They are not real user search logs. Many of the cited StatWharf pages belong to one cluster of SAP governance categories, so results lean on that cluster.

Citations by engine

On September 12, 2026, the same 100 prompts ran on three engines on the same day.

The Perplexity check is small. It used 11 prompts on day 1 and 10 on day 2, because the engine stopped accepting searches after 10. It shows that Perplexity can cite these pages. It does not give a reliable rate.

Google placed StatWharf in its ordinary results more often than in its AI Overview. Seven of the 8 organic appearances were SAP categories.

Earlier logged-out and signed-in runs

The first runs took place on September 8 and 9, 2026. ChatGPT ran logged out. Google ran in a signed-in browser with unknown personalization.

The logged-out ChatGPT rate was 10 of 100 across both arms. The signed-in run three days later was 32 of 100. Three things changed between the two runs: account state, model routing and four more days of crawling. The data cannot say which one moved the rate.

Gemini answered 58 of 100 prompts with no link at all. Most of the links it did show were shopping links, not editorial sources. The Google AI Overview capture on September 9 was incomplete. Only 47 of 100 records hold a verified AI Overview, and secondary source panels were not all expanded. The zero therefore covers the captured links only.

What ChatGPT searched

ChatGPT exposed its searches in the live response stream. Gemini and Google ran their searches on the server and sent none to the browser. These figures come from the September 12 run. The fan-out study gives the full method.

ChatGPT kept the category phrase and added vendor names and a year. Searches that led to a StatWharf citation kept the phrase at the same rate as searches that did not. The phrase is a property of how ChatGPT expands these prompts. This data does not show that it predicts a citation.

Which vendors the answer named

StatWharf checked the answers that cited a StatWharf page against the vendors listed on that page. This pass used 105 ChatGPT runs from September 12 to September 15, 2026, with 987 vendor rows in total. A vendor counts as named only with its full name in the answer text.

Vendors that ChatGPT searched for were named about twice as often. The direction is not established. ChatGPT writes its searches before it reads any page, so a vendor name in a search mostly reflects the model’s own prior knowledge. Runs within a category are not independent, so any interval on these rates is optimistic.

Prompt wording

The phrasing panel ran 11 categories in five wordings on two days. The wordings were top {category}, which {category} for a 15-person team, alternatives to {vendor} {category}, {category} for startups and {category} vs {vendor}.

The gaps between wordings are not reliable. Every wording rose from day 1 to day 2, by a mean of 2 prompts out of 11. The largest gap between wordings was 27 percentage points. A power check found that 33 prompts per wording detect only a gap of about 32 points. A third day was therefore not run.

The persona runs asked the same category prompts with a short description of the user in front. These samples are also small.

A neutral prefix with no persona lost about as many citations as the persona prefixes. Prompt length, or any extra context before the category phrase, explains much of the drop. The persona alone does not.

How stable a citation is

The same prompt does not always give the same result. Two measures show the size of this noise.

The SAP panel held at 16 of 20 on all three days. Across all repeated prompts, about 1 group in 4 flipped between cited and not cited at least once. A test of a page change therefore needs far more runs than a single day of a small panel provides.

FAQ

How often did ChatGPT cite StatWharf?

In the September 12, 2026 run, ChatGPT cited StatWharf in 32 of 100 answers. In 30 of those 32, the cited page was the page for the category in the prompt. A logged-out run on September 8 cited StatWharf in 10 of 100 answers.

Did Gemini or Google AI Overview cite StatWharf?

No. Gemini cited StatWharf in 0 of 100 answers on September 9 and again on September 12. Google AI Overview cited it in 0 of 100 on September 12. Google did show StatWharf in ordinary organic results on 8 of those 100 result pages.

Does ChatGPT search the prompt exactly as typed?

No. None of the 211 searches in the September 12 run matched the prompt. ChatGPT kept the category phrase in 68% of searches and added a year in 76%. More detail is in the fan-out study.

Does prompt wording change the citation?

The results moved by wording, but day-to-day noise was about as large as the gaps. The “top” wording cited the intended page in 18 of 22 runs and the “vs” wording in 12 of 22. The sample cannot separate those gaps from noise.

Why are the samples so small?

Each run is a live browser session, and engines limit how many prompts one account can run in a day. Small samples are stated next to each figure. Read figures with fewer than about 20 runs as examples, not rates.

Questions about the method can go to editorial@statwharf.com.

Factual corrections are free: editorial@statwharf.com.