AI citations vs search results: how much do they overlap?

A cited page is one URL an engine cites on one prompt-day. Coverage is the share of cited pages that also appear in the top 10 results of at least one of the five search APIs for the same prompt, searched within three days.

ChatGPT Search cited 4,184 pages on matched prompt-days. 18.0% of them appeared in a search top 10. Google AI Mode cited 2,234 pages, and 36.9% appeared. So roughly 4 in 5 pages ChatGPT Search cites are not in any of the five top-10 lists, and about 2 in 3 for Google AI Mode.

A cited site is the registrable domain of a cited page, such as wired.com. At the site level, 44.8% of ChatGPT Search's cited sites and 56.1% of Google AI Mode's appeared in a same-prompt search top 10.

Widening the search does not close the page gap. Across every result any API returned for the same prompt over the whole window, page coverage rises only to 26.4% for ChatGPT Search and 50.1% for Google AI Mode. Across all 124,060 results for every prompt, it reaches 27.4% and 52.2%.

The engines often cite a site that search also ranks, but a different page on it.

The inverse view starts from search. Of the sites in the five APIs' combined top-10 lists for a matched prompt-day, 94.9% were not cited by ChatGPT Search that day, and 94.4% were not cited by Google AI Mode.

Looking across all three weeks for the same prompt, 89.9% of those sites were never cited by ChatGPT Search for that prompt, and 87.4% never by Google AI Mode. Part of this gap is arithmetic. The engines cite 3.8 (ChatGPT Search) and 3.3 (Google AI Mode) sites per prompt-day, while the five top-10 lists together hold about 33 distinct sites.

Each API on its own sees a small slice. For ChatGPT Search, single-API page coverage ranges from 3.3% (Linkup) to 7.7% (You.com). For Google AI Mode, it ranges from 6.0% (Exa) to 21.1% (Tavily).

Tavily and You.com match Google AI Mode best, at about one in five cited pages each. No API stands out for ChatGPT Search: You.com, Exa, and Parallel each find 6% to 8% of its cited pages. Combining all five more than doubles the best single API for both engines.

Site overlap here is the Jaccard overlap of two APIs' top-10 site lists for the same prompt on the same day: shared sites divided by all sites in either list. Across the 10 API pairs, average site overlap ranges from 4.1% (Exa and Linkup) to 19.4% (Tavily and You.com).

Each API is far more consistent with itself. Asked the same prompt on its next scheduled day, each API's site overlap with its own previous list ranges from 54% to 87%. The APIs are stable, and they disagree with each other. That is why combining them finds more of what the engines cite.

Engines change their citations from run to run, so a perfect match is not the right bar. When ChatGPT Search answered the same prompt again within a week, 43.6% of the sites it cited on one day reappeared. The five APIs' same-prompt top 10 covered 44.8%, about the same share.

For Google AI Mode, search covers more than the engine's own repeat run: 56.1% against 37.2%. The two engines agree with each other far less. Only 12.7% of the sites ChatGPT Search cited were also cited by Google AI Mode for the same prompt on the same day.

A named-brand site is the primary website of a brand the same answer names. 49.8% of the sites ChatGPT Search cited were named-brand sites. In the five APIs' top-10 lists for the same prompts, named-brand sites were 7.3% of results. That is 6.8 times the search rate.

Google AI Mode cites named-brand sites less: 30.9% of its cited sites, against 8.9% in search, or 3.5 times. When search did return a named brand's site, ChatGPT Search cited it 57% of the time and Google AI Mode 34% of the time.

This is the clearest choice the engines make that search does not: they go to the brand's own site to back the brand they recommend.

The citation-to-search ratio compares a site's share of engine citations with its share of search results on the same matched prompt-days. A ratio of 1 means the engines cite the site as often as search returns it. Below 1 means search returns it more. We add one to each count so a site with zero citations still gets a ratio.

Tracxn appeared in search results for 70 prompts and was cited on 1 prompt-day. RocketReach appeared for 43 prompts and was never cited. The list is company-profile databases (Tracxn, RocketReach, CB Insights), job and map listings (Indeed, MapQuest), market-data pages (Stock Titan, Investing.com), a statistics compilation site (WorldMetrics), and general news (Business Insider, CNN, CNBC, Yahoo).

These sites have a citation-to-search ratio above 1: the engines cite them more often than the five APIs return them. Among sites cited on at least 10 matched prompt-days, the top ratios belong to the FTC (8.1), Lowe's, The Home Depot, TechRadar, and Square.

Gartner was cited on 33 prompt-days and Wired on 31, the most in this list after YouTube. Several vendors (Square, Stripe, Airwallex, Shopify, Microsoft, Nvidia) are cited for their own product pages more than search ranks those pages for the same prompts.

Search returned YouTube for 121 of the prompts ChatGPT Search answered, and ChatGPT Search cited it on 1 prompt-day. Search returned Reddit for 98 prompts, and ChatGPT Search cited it on 13 prompt-days, all within 3 prompts.

Google AI Mode does the opposite. It cited YouTube on 113 prompt-days across 67 prompts, a citation-to-search ratio of 6.97. It cited Reddit on 87 prompt-days, a ratio of 3.37. Both engines saw the same search supply. The difference is in what each engine chose to cite.

We checked every search result against all citations in the public Parse index, across all engines and dates. 5.8% of all 124,060 results came from a site that no AI answer has ever cited.

The share varies by API. 13.3% of Exa's results and 9.8% of Linkup's came from never-cited sites. For Tavily and You.com the share is under 1%. The two APIs whose results look most like the cited web are also the two that best match Google AI Mode's citations.

We grouped prompts by the industry group of their primary market and kept groups with at least 5 prompts. ChatGPT Search's site coverage ranges from 36.8% in Information Technology to 65.6% in Food and Beverage.

No industry group gets above three in four. Groups hold 5 to 12 prompts each, so differences of under 10 points between groups are within what a few prompts can move.

ChatGPT Search's site coverage was 44.1%, 45.7%, and 43.8% in the three full weeks. Over the same weeks, its average searches per answer rose from 1.01 to 2.32, and reached 2.95 in the first days of October.

Running more searches did not move the share of cited sites that plain search also returns. The October week holds only 19 matched prompt-days, so its 51.5% is not a trend.

The panel is 214 fixed buyer prompts. Each engine answer was paired with each search API's list for the same prompt on the nearest day within three days (average gap 0.60 days). A prompt-day counts only when all five APIs had a list, and only when the engine cited at least one source. That excluded 135 of 783 Google AI Mode answers and 16 of 821 ChatGPT Search answers with no citations.

Pages are compared as host plus path, lowercased host, without www, query strings, fragments, or trailing slashes. Sites are registrable domains. Each search list is the API's top 10 at its default fast or standard setting; a deeper list would raise coverage.

Google AI Mode answers in this window carry fewer citations than earlier in 2026. On August 26, its answers averaged 9.48 citations; from August 29 they averaged about 3.73, and links in the answer text fell at the same time. We could not tell whether the engine or the way answers are collected changed, so Google AI Mode figures here describe the shorter citation lists only.

Parse's own pages appeared in search results for 49 prompts and were cited on 18 prompt-days. We left them in the counts. The window is three weeks; the result is a snapshot, not a trend.

Ranking in web search helps, but it does not decide the citation. About half of the sites the engines cite also rank for the same prompt, but most cited pages do not. Optimize the page the engine wants, not only the site that ranks.

Make your own site citable for the claims an answer makes about you. ChatGPT Search backs named brands with their own sites in half of its cited sites. Product, pricing, and documentation pages that state facts plainly are the pages it cites.

Split the work by engine. For Google AI Mode, YouTube and Reddit presence matters more than search rank suggests. For ChatGPT Search, it matters less. Do not rely on company-profile databases or statistics compilation pages to represent your brand. Search returns them, and the engines rarely cite them.