Do shared sources lead to the same AI recommendation?

A matched pair is one answer from each engine to the same organic prompt in the same run. ChatGPT Search and Google AI Mode chose different top brands in 5,267 of 9,854 matched answer pairs that shared at least one cited page, or 53.4504%. They chose the same top brand in 4,587 pairs, or 46.5496%.

A shared page does not determine the first recommendation. Audit the cited page and the selected brand separately instead of treating shared evidence as a shared outcome. BrightEdge found that three AI platforms returned different brand recommendations for 61.9% of queries. This study narrows that question to two engines, one first recommendation per answer, and pairs that already share an exact page.

The engines chose the same top brand in 4,587 of 9,854 pairs that shared an exact page, or 46.5496%. The rate was 39.6622% when the pair shared a website but no exact page and 31.8330% when it shared no website. Exact-page overlap was associated with 14.7166 percentage points more agreement than no website overlap.

Source overlap is useful context, not proof of a ranking mechanism. The comparison is descriptive and does not show that a shared page caused either recommendation. Ahrefs found only 13.7% cited-URL overlap between Google AI Overviews and Google AI Mode while their answers stayed similar in meaning. Which sources an engine picks and which brand it puts first can move independently.

Among the 5,267 shared-page pairs with different top brands, at least one answer still named the other engine's top pick in 4,022 pairs, or 76.3623%. Both answers named the other top pick in 1,783 pairs, one answer did so in 2,239, and neither did so in 1,245.

The engines often considered overlapping candidates and changed only which brand they put first. Track the full shortlist when diagnosing a first-place difference. Semrush's ghost-citation study showed that a citation and a brand mention are separate events; this result adds recommendation rank as a third field that should not be collapsed into either one.

The engines chose the same top brand in 4,036 of 8,765 pairs that shared one page, or 46.0468%. Agreement reached 50.8097% when they shared two pages and 48.5149% when they shared three or more. Those groups contained 988 and 101 pairs.

A longer list of common pages did not produce a reliable shared winner in this observed cut. The groups with multiple shared pages are smaller, so use them to guide review rather than as a cutoff that proves cause.

Among industries with at least 20 shared-page pairs, Consumer Electronics chose the same top brand in 49 of 68 pairs, or 72.0588%. Software reached 53.5070%, Financial Services 45.4039%, Sales and Marketing 39.7436%, Internet Services 35.2941%, and Payments 27.5862%.

Use an industry baseline before treating a brand's result as unusual. These rates describe the observed prompt mix and do not show that industry caused the recommendation difference.

ChatGPT Search put Upwork first while Google AI Mode put Toptal first in 39 shared-page pairs across six prompts. Growth Marketing Pro and Soar followed with 18 pairs, Global Forest Watch and Airbus with 17, Grafana and SigNoz with 16, and Zyte and Bright Data with 15.

A repeated brand pair identifies where to inspect the prompt, page, and ranking language together. It does not show that the shared page endorsed either brand or caused the order.

InterTeam Marketing's Reddit agency list was the most repeated shared page in different-winner pairs, appearing 55 times across seven prompts and 25 top brands. Sheltera's utility vegetation management page appeared in 46 pairs, a Reddit discussion about Quora marketing agencies in 29, KYRO AI's utility vegetation software guide in 28, and the MentorCruise homepage in 28.

The leading pages are lists, service pages, and discussions that can support several options. Audit how each engine uses the page inside the answer instead of treating the citation as an endorsement of the first brand.

The main result, comparing page addresses in their cleaned form, was 5,267 of 9,854 pairs, or 53.4504%. Comparing the addresses exactly as each engine recorded them produced 855 of 1,658, or 51.5682%. The June 1 through July 16 cut produced 5,266 of 9,853, or 53.4457%, and taking only the first qualifying run per prompt produced 1,100 of 2,111, or 52.1080%.

The window contained 557,328 answers. We retained 215,350 with one unambiguous first recommendation, formed 81,796 same-prompt answer pairs, and required a cited page on both engines to retain 64,767 pairs. Duplicate page addresses were merged after fragments, trailing slashes, and eight standard tracking parameters were removed. The study does not show that a cited page supports a top brand, caused the ranking, or made the recommendation accurate.

Shared-page pairs chose different top brands 53.4504% of the time. Agreement rose from 31.8330% with no shared website to 46.5496% with a shared page, while 76.3623% of shared-page disagreements still named the other engine's top pick.

Track the first recommendation, full shortlist, registered website, and cleaned cited-page address for each priority prompt and engine. Review repeated brand-pair differences in the answer text before changing content. Rerun this method over the same window length next quarter before treating a difference as real movement.