TL;DR
- Google Search and the OpenAI Responses API are both partial proxies for ChatGPT Web. Together they recover about a quarter of the exact pages ChatGPT cites, and about half of the cited hostnames.
- The API’s own citations are the most precise proxy: 23% of them are pages ChatGPT Web also cites on GPT-5.5, 41% at the hostname level. The API’s candidate list has the best recall. Google alone is the weakest on both.
- Proxies agree with ChatGPT Web far better on hostnames and entities than on exact URLs. Track sites and names, not only pages.
Few teams can trace ChatGPT Web directly. Most visibility tools ask the OpenAI API instead, or infer from Google rankings. Part 2 showed what ChatGPT Web cites; this part asks how much of that the two common proxies can see. The data comes from Understanding ChatGPT Citations.
How the proxies were collected
For each of the 80 questions we collected three things within the same three-day window as the ChatGPT Web runs. Google Search: five pages of organic results through a search data provider, 43.2 results per question on average. The Responses API with web search on, 20 calls per question on gpt-5.3-chat-latest and gpt-5.5, which yields two lists: the candidate pages the API retrieved (on average 27.9 and 25.7 per call) and the pages it cited in its own answer (5.8 and 5.4). We compared each list with what ChatGPT Web cited across its 20 runs, as precision, recall and F1, at three levels: exact URL, hostname and entity.
Finding 1: exact pages are hard to recover, hostnames much easier
| Proxy, GPT-5.5 | URL precision | URL recall | Hostname precision | Hostname recall |
|---|---|---|---|---|
| Google Search | 4.1% | 6.4% | 12.5% | 22.8% |
| API candidates | 7.1% | 22.5% | 17.1% | 38.1% |
| API citations | 23.3% | 13.0% | 40.5% | 23.9% |
| Google and API candidates together | 5.7% | 26.2% | 12.8% | 48.7% |
Google and the API candidates combined recovered 26.2% of the exact URLs ChatGPT Web cited on 5.5 (28.3% on 5.3), but 48.7% of the cited hostnames (50.3% on 5.3). A proxy that misses the page often still finds the site.
Finding 2: the API’s citations are the most precise proxy, and got more precise
The pages the Responses API cited in its own answers overlapped most with ChatGPT Web. Their URL precision rose from 16.4% on 5.3 to 23.3% on 5.5, and hostname precision from 27.7% to 40.5%; both increases held after correction. Recall barely moved. Ranking quality told the same story: NDCG@10 for API citations rose from 0.203 to 0.246 at the URL level and from 0.352 to 0.424 at the hostname level. The API’s candidate list, by contrast, had the best recall (22.5% of URLs, 38.1% of hostnames) at low precision. Google alone was weakest on every measure.
Finding 3: at the entity level, no proxy dominates
Entities include brands, products and organizations named in cited pages. On 5.5 the API citations had the highest precision (30.1%) and the API candidates the highest recall (38.0%); Google and API candidates together reached 45.1% recall. On 5.3 no proxy was consistently best. The entity layer mixes brands with other named things, so we do not draw separate conclusions for brands.
What this says for GEO
If you measure ChatGPT through the API, you are seeing a related but different system, and the gap is largest at the page level. Hostname and entity presence are the measures the proxies agree on, so a report that says “your site and your name appear” is on firmer ground than one that says “this page is cited.” And the newer model made the API a better proxy, which means the quality of any proxy has to be re-checked when the model changes. Source asks ChatGPT through more than one path for this reason and keeps the answers apart.
What this part does not show
The API runs used gpt-5.5 with reasoning turned off and no system prompt, which is one configuration among many. Google results came through a provider, not a browser. The proxies and ChatGPT Web were collected within three days of each other, not simultaneously. None of this measures whether a page that a proxy shows will be cited by ChatGPT Web in future.
Continue with Part 4: can citations be predicted from outside?