Do ChatGPT and Google cite the same page?

A shared-website pair is one question where both ChatGPT Search and Google AI Mode cited the same website. In 53,032 such pairs, the two engines landed on at least one identical page 19,323 times, or 36.4%. In the remaining 33,709 pairs, or 63.6%, both engines cited the site and neither picked the same page.

A domain-level citation report overstates how much a page-level win carries between engines. Audit which address each engine cites, not only which site. This measures whether the page is the same, not whether it is better, and it does not show that either engine made the stronger choice.

Among the 21 websites with at least 100 shared-website pairs, TechRadar led at 86.5%, or 109 of 126. Consumer Reports reached 68.3%, Healthline 67.0%, Forbes 66.2%, and RTINGS 65.7%. At the other end, monday.com agreed on the page in 7.2% of its 111 pairs, ClickUp in 8.8% of 137, Atlassian in 9.3% of 118, LinkedIn in 11.6% of 146, and Salesforce in 11.8% of 119.

The 36.4% average hides a twelvefold spread across websites. Benchmark against similar publishers, because agreement differs by site type.

Grouped by publisher type, health sites agreed on the page in 60.9% of 647 pairs, reference sites in 55.8% of 1,408, government sites in 55.1% of 514, and editorial sites in 54.1% of 1,722. Social platforms came last at 18.5% of 9,292 pairs, below software vendor sites at 36.5% of 13,261.

Where a publisher keeps one clear article per topic, both engines find it. Where a site holds many near-equivalent pages, the engines split between them. Expect page-level agreement to track how consolidated a site's content is, which is a property a site owner can change.

Of 225,640 ChatGPT Search page citations, 58,610, or 26.0%, pointed at a website's root address. Of 992,342 Google AI Mode page citations, 10,794, or 1.1%, did. That is a twenty-four-fold difference over the same 33 days on the same questions.

This is the mechanism behind the leaderboard. The engines are not weighing the same pages and picking differently; they are citing different kinds of page. A homepage citation and an article citation can never match, so any site whose citations split this way records a low agreement rate automatically, not because the engines disagreed.

On ClickUp, 46.2% of ChatGPT Search's 327 citations pointed at the homepage, against 0% of Google AI Mode's 1,263. monday.com ran 52.0% against 0%, Rippling 50.4% against 0.14%, and Salesforce 24.8% against 0%. On the publishers that agree most, both engines sat at 0%: TechRadar, Forbes, Healthline, Consumer Reports, and RTINGS.

For a vendor site these are two separate pieces of work. The homepage is what earns a ChatGPT Search citation, and a specific article is what earns a Google AI Mode citation. A single combined count of a brand's AI citations merges two different pages doing two different jobs.

Google AI Mode cited an average of 54.5 distinct pages and 39.5 distinct websites per question. ChatGPT Search cited 12.6 pages and 9.7 websites.

This makes the low agreement rates conservative rather than inflated. The broader engine had more chances to land on the same page by coincidence and still did not, so the disagreement is not a side effect of one engine citing too few pages to overlap.

Across 36 industries with at least 200 shared-website pairs, Health Care led at 49.3% of 1,409 pairs, ahead of Agriculture and Farming at 48.3% of 211 and Gaming at 46.5% of 409. The lowest were Artificial Intelligence at 29.2% of 1,684 pairs, Information Technology at 29.4% of 2,647, and Collaboration at 30.3% of 281.

Categories served by established reference publishers agree most, and categories served mainly by vendor marketing sites agree least. Expect page-level agreement to fall in a market as its evidence base shifts from independent publishers toward vendor-owned content.

Addresses are compared only after they are cleaned to a standard form, because 71.8% of ChatGPT Search citation addresses carry a tracking parameter that 0% of Google AI Mode addresses carry. Reddit threads are reduced to the thread identifier because the two engines attach different title text to the same thread. One research archive is excluded because the engines cite the same document under two different address forms. Parse's own website and its agency partner's website are excluded because they are our own. Including every excluded website moves the headline from 36.4% to 36.8%. Restricting to the 19,899 pairs where each engine cited exactly one page on the website gives 34.9%.

The conclusion survives its own sensitivity checks, and the strict one-page-each rule removes any advantage the broader engine gains from citing more pages. The result describes page identity over one 33-day window on two engines. It does not describe other engines, other windows, or whether either page was more accurate.

The two engines shared no page in 63.6% of 53,032 shared-website pairs. Agreement ranged from 86.5% to 7.2% by website, and ChatGPT Search cited homepages at 26.0% against Google AI Mode's 1.1%.

Report AI citations at the page address for each engine separately, not as one website-level count. Treat the homepage and the article library as two assets serving two engines. On sites carrying many near-duplicate pages, consolidate to one clear page per buyer question so both engines have a single place to land. Recheck on a fixed schedule, because a page-level result in one window is not a durable position.