Measuring GEO Visibility: Why Google Finds Only a Quarter of ChatGPT's Sources

In short: Google results and the OpenAI API together find only about a quarter of the pages ChatGPT names as sources. At hostname level it is about half. For us that means measuring sources over many runs, counting page and hostname separately, and recording the model for each answer. Our Umami data shows 29 visits from AI assistants in 90 days. We have not yet done the test runs themselves.

In the GEO post from 14 September we checked whether AI systems are allowed to read our pages and can understand them. What that post lacks is a measurement of whether they actually name us. Since October there is a study that explains why the obvious shortcuts do not work.

Measuring GEO visibility means finding out how often a page or a hostname appears as a source in AI answers. Because ChatGPT rewrites the question, retrieves different pages on every run and names only some of them, the measurement needs repeated runs, an evaluation at both page and hostname level, and a record of which model produced each answer.

What the study measured

Yunxuan Jiang and four co-authors observed ChatGPT in its web interface. They took 80 questions from the Search Arena dataset, half English and half Chinese, and asked each question 20 times. The runs took place between 29 March and 29 July 2026 for GPT-5.3 Instant and GPT-5.5 Instant. The authors then checked how many of the cited sources can be reproduced from Google results (via SerpAPI) and from the OpenAI Responses API. The paper is an arXiv preprint and was not peer reviewed.

Unit: share in percent. Measurement setup: ChatGPT Web, 80 questions times 20 runs, 29 March to 29 July 2026. These are the study's figures; we did not measure them ourselves. We cross-read these values in the PDF. The table with the single comparisons (Google alone, API alone) did not extract cleanly from the text, so we take those figures only from Search Engine Journal of 18 October, which reports that Google alone covered 9.4 and 6.4 percent of the cited pages.

Three findings matter more in practice than the percentages:

• In the visible searches, ChatGPT never used the question in its original wording. It rewrites it, and for Chinese questions a large share of the searches ran in English only (67.0 and 69.6 percent according to Search Engine Journal).
• Sources that are retrieved more often across runs are also cited more often. One run therefore shows chance more than pattern.
• Two sessions silently switched to a fallback model in the middle of a measurement series, after 14 and after 16 successful runs. Anyone who measures repeatedly has to check for each answer which model produced it.

There is also the gap between retrieval and citation. A page can appear in the search results and still go unnamed. The authors advise looking at visibility in retrieval and visibility in citation separately, and evaluating at hostname level in addition to the exact address.

What Umami counts on our site

The only data source we could read without running a new measurement is our web analytics. Umami 3.3.1 runs on the dev VM 192.168.181.3 and counts senn-tech.com, as described in our switch from Matomo to Umami. With a read-only SQL query on website_event, we counted the visits of the last 90 days (8 August to 5 November 2026) by referrer domain and utm_source. In our data, links from ChatGPT often carry utm_source=chatgpt.com; we saw that in the query. We did not find a statement from OpenAI on this that we could cite here.

Unit: visits (visit_id), measured on 5 November 2026 with Umami 3.3.1, Postgres database, site senn-tech.com. The bot filter is the procedure from our Umami report (known headless screen sizes, and visits from SG, CN, HK, RU without a referrer, are removed). All 29 AI visits fall into the filtered share, which is about 1.9 percent of the filtered visits. For comparison, google.com alone is the top referrer with 474 visits.

The visits spread over 22 different paths. The three pages that can be attributed with certainty through the chatgpt.com referrer are measurement reports: the self-hosted operation of Qwen3.8-27B, the FP8 KV test of Qwen3.8-Flash-Next and the Windows comparison of coding agents. None of this tells us how often ChatGPT names us. Someone who sees the source and does not click never shows up in Umami, and a click without a referrer and without utm_source counts as a direct visit.

Still to measure: repeated runs in ChatGPT

We have not done the runs. The procedure is fixed; there are no results yet.

1. Ten questions on topics we have written about (for example local models on RTX 5090, Proxmox backup, mail gateway operation), each in German and in English.
2. Each question 20 times in ChatGPT on the web, every run in a new conversation, without history, in the default model.
3. Per run, record the displayed model, every address named, and the date. A run on a different model is counted separately.
4. Evaluate at two levels. Page level: in how many of the 20 runs is exactly this address named. Hostname level: in how many runs does senn-tech.com appear anywhere.
5. Repeat after four weeks and after every model change.

It is open whether 20 runs are enough for our topics. The study used 20 for 80 questions in English and Chinese and did not examine German questions.

Still to measure: AI crawler requests

Whether OAI-SearchBot and ChatGPT-User fetch our pages has not been evaluated. Our robots.txt names both. In its crawler documentation, OpenAI describes that OAI-SearchBot is set for search and GPTBot for training, each independently of the other. At the Caddy edge 192.168.181.1, where our notes place the access logs, the newest log file is from 18 August. For November the count therefore cannot be read there at the moment. That has to be clarified before anyone quotes a crawler number.

What is established and what is open

Established, from the study: Google plus API reproduce about a quarter of the cited pages and about half of the cited hostnames. Any one run is incomplete. The model can change during a measurement series.

Established, from our data: 29 visits from AI assistants in 90 days, 25 of them from ChatGPT, out of roughly 1,520 filtered visits.

Open: whether the results hold for German questions and for models from GPT-5.6 on. Search Engine Journal points out that the study did not test any commercial tracking tools. Also open is how often ChatGPT names us. The repeated measurements are missing for that.

Sources last checked: 2026-11-05.

A 10-minute check

Open your web analytics and filter the visits of the last 90 days by referrer chatgpt.com and by utm_source=chatgpt.com. Write down the number and the landing pages. Then ask ChatGPT the same question about your most important topic five times, each in a new conversation, and note whether your domain appears in every run, in none, or in only some. If you want help setting up a measurement like this, our offer is on the SEO and GEO page, and the wider context sits in the topic area AI explained.

Further Reading

• Understanding ChatGPT Citations: Evidence from Production Web Search Across Model Generations, Jiang et al., arXiv preprint of 11 October 2026, without peer review; retrieved 5 November 2026.
• OpenAI's API Only Partly Matched ChatGPT's Citations, Paper Finds, Search Engine Journal, 18 October 2026; source for the single values of Google alone and the language shares.
• Overview of OpenAI Crawlers, OpenAI, on OAI-SearchBot and GPTBot.