Do the pages AI cites actually support what AI says?

A claim-citation pair is one automated reading of what an AI answer says about one brand, matched to one page cited for that brand in the same answer, where an automated reading of the cited page's own coverage of that brand completed. A praise pair is a claim-citation pair where the AI claim is positive. A praise pair is backed when the cited page's coverage of the brand also reads positive.

Of 119,869 checked praise pairs, 71,377, or 59.5%, were backed. The other 48,492, or 40.5%, were not. That is about 2 in 5 praise citations pointing to a page that does not read positive about the brand the answer praised.

Among the 119,869 checked praise pairs, 45,390, or 37.9%, cited a page whose coverage of the brand was neutral. Only 2,224, or 1.9%, cited a page that was negative about the brand, and 878, or 0.73%, cited a page with mixed or hedged coverage.

AI praise citations almost never point to a page that disagrees. They point to a page that names the brand and takes no stance, such as a docs page, a directory row, or a feature list. The praise comes from somewhere else; the citation supplies presence, not endorsement.

On ChatGPT Search, 21,892 of 31,773 checked praise pairs, or 68.9%, were backed. On Google AI Mode, 49,485 of 88,096, or 56.2%, were backed. The gap is 12.7 points.

Read the other way: 43.8% of Google AI Mode praise pairs were not backed, against 31.1% on ChatGPT Search. This comparison describes what we observed and does not identify the cause.

Support has a weaker test that runs first: does the cited page contain the brand at all. On the same window, that check decided 295,727 brand-citation pairs. The page named the brand in 176,948, or 59.8%, and did not in 118,779, or 40.2%. Another 131,264 pairs could not be decided because the page was not fetched yet, the fetch failed, or no stored copy existed.

Our earlier report measured this brand-presence gap at 38.5% on a smaller window. The claim-direction result in this report applies after a page passes the brand-presence test, so the two gaps combine: many cited pages never name the brand, and among pages that do, 2 in 5 do not back the praise.

Documentation pages backed 375 of 980 checked praise pairs, or 38.3%. Social pages backed 45.4%, finance pages 48.7%, health pages 51.2%, and education pages 51.3%. At the other end, SaaS sites backed 64.5%, reference pages 63.7%, and news pages 63.5%.

The pattern follows how each page type writes. Documentation states facts without judgment, so it reads positive less often. Vendor and review pages carry stance. A citation's source type is a rough predictor of whether it will back praise.

Among domains with at least 400 checked praise pairs, linkedin.com backed 567 of 1,880, or 30.2%, and newrelic.com backed 30.2%. Reddit backed 50.6% and Medium 52.5%. At the top, linear.app backed 98.9% and aidash.com 90.7%.

The extremes are mostly brand-owned domains, which are cited for their own brand and read positive about it. The low end is profile and documentation hosting, where a page names a brand without praising it.

When the cited page lived on the brand's own site, 35,343 of 54,320 checked praise pairs, or 65.1%, were backed. On third-party pages, 36,033 of 65,548, or 55.0%, were backed.

Even on the brand's own site, about 1 in 3 cited pages did not read positive. Pages like documentation, changelogs, and pricing tables state facts without making the case for the product, and they earn many of these citations.

Brands in Community and Lifestyle had praise backed in 46.9% of checked pairs and brands in Clothing and Apparel in 79.6%, a spread of 32.8 points across the 26 industry groups with at least 1,000 pairs. Software sat at 56.7%, Financial Services at 54.0%, and Consumer Goods at 70.5%.

Consumer categories are covered by review and commerce pages that take a stance. Software and finance brands are covered by documentation and comparison tables that do not. Compare your brand with its industry rate, not the 59.5% overall rate.

The checked set contained 817 criticism pairs, where the AI claim about the brand was negative. The cited page was also negative about the brand in 64 of them, or 7.8%. It was positive about the brand in 381, or 46.6%, and neutral in 358, or 43.8%.

So a page cited next to criticism is more likely to praise the brand than to criticize it. The sample is small relative to the praise set, because AI answers rarely attach criticism to a brand-attributed citation, and the share should be read with that in mind.

Of 150,811 checked claim-citation pairs, 119,869, or 79.5%, carried a positive AI claim. Neutral claims supplied 17.0%, hedged 1.8%, mixed 1.2%, and negative 0.54%.

This is why the headline conditions on praise pairs. What AI says about brands next to citations is overwhelmingly positive, so the support question is in practice a praise-support question.

The same cited page and brand can appear in many answers. Counting each page-brand pair once, using its latest observation, leaves 54,450 unique pairs, of which 29,828, or 54.8%, were backed.

The unique-pair rate is 4.8 points lower than the 59.5% pair-level rate, which means repeatedly cited pages are somewhat more likely to back praise. The headline is not an artifact of a few heavily repeated pages.

Across the eight full weeks in the window, the backed share of checked praise pairs ranged from 57.8% to 60.8% with no trend, while weekly volume grew from 2,186 to 30,824 pairs.

The rate is a stable property of how these engines cite, not a one-week event. A partial first week of 24 pairs is excluded from the chart.

A pair enters the checked set only when the page was cited for that brand in the answer, the page named the brand, and an automated reading of the page's coverage of that brand completed. Within the window, 150,811 of 151,169 brand-attributed pairs recorded in the window, or 99.8%, had a completed page reading. Backed is a direction-level match between two automated readings, not sentence-level fact-checking: a neutral page can still contain the specific fact the answer used, and a positive page can praise the brand for a different reason than the answer gives.

Pages that failed the brand-presence test, pages that could not be fetched, and citations not attributed to a brand are outside the checked set. The window ends July 19, 2026, where this analysis layer thins, and covers ChatGPT Search and Google AI Mode only. The headline was recomputed two independent ways, on unique page-brand pairs and through a second join path restricted to validated readings, with consistent results.

AI praise citations fail to back the praise 40.5% of the time, the gap is wider on Google AI Mode, and the pages that fail are mostly neutral pages that name the brand without taking a stance.

Put stance where AI already cites you. Find the pages AI attaches to your brand, and add explicit evaluative statements, outcomes, and comparisons to the ones that only state facts. Documentation and profile pages earn the citation; a page that also makes the case for the product turns the citation into support. Then rerun the same window next quarter and watch the backed rate, not just the citation count.