When Google Halved AI Mode Sources, Brand Sites Won Share but Lost Answers

GEOforge Research · Citation Data

When Google cut the number of sources in its AI Mode answers by more than half, brand websites kept a bigger slice of a much shorter list. They also appeared in fewer answers. Both halves are true, and most dashboards show only one.

By the GEOforge TeamGEOforge7 min read

In late August 2026 Google AI Mode started attaching about half as many sources to each answer. We wanted to know who paid for that. Did brands’ own websites lose their citation slots along with everyone else, or did third-party sites take most of the cut? The answer has two sides, and a report that shows only the first would flatter your site.

What is the short answer?

• A brand’s own site took a bigger share of a smaller list. Across seven brand sites, own-site share of AI Mode citations rose from 2.7% to 4.4% on average. It rose in 7 of 7 sites (barely, in one).
• But the site appeared in fewer answers. The share of answers citing the brand’s own site at least once fell from 33% to 27% on average, and fell in 6 of 7 sites.
• Third-party sources took the larger cut. Per answer, own-site citations fell 36%. Everyone else’s fell 57%. That is why the share rose even though presence fell.
• It is specific to the engine whose lists shrank. In ChatGPT, where lists stayed at about 8 sources, own-site share did not rise (3.2% to 3.1%).

What did we measure?

GEOforge runs each brand’s tracked prompts against ChatGPT, Google AI Mode and Google AI Overviews on a schedule and stores the sources each answer attaches. For this piece we took seven brand sites (six brands; two of the sites are regional versions of the same brand) that have data on both sides of the change (one of them has only a single full week before it; we show results with and without it). For each site we counted how many of the sources in each answer sit on the brand’s own domain, including its subdomains.

“Before” is 6 July to 16 August 2026. “After” is 31 August to 4 October. We left out 17–30 August so the change itself is not averaged into either side. We compared only prompts that were measured on both sides, which gives 53,274 AI Mode answers. No prompt in that set names the brand, so these are generic, unbranded questions. Every figure is a brand-by-brand average, so a site with more answers does not outweigh a site with fewer.

Did brand sites keep their citation slots?

Their share of them went up. In five of the seven sites, AI Mode’s source lists went from about 22 per answer to about 8. In all five, the brand’s own share of citations rose. The two regional sites started with shorter lists, about 9, and fell less, to about 7. Their share barely moved in one and rose in the other.

Did they appear in more or fewer answers?

Fewer. Share of citations is a ratio, and a ratio can rise while the numerator falls. Per answer, a brand’s own site went from 0.55 citations to 0.35, a drop of 36%. Third-party sources went from 17.5 per answer to 7.5, a drop of 57%. The brand site lost slots too, just fewer of them. The share of answers that cite the site at all fell from 33% to 27%.

Is this just our clients’ content work paying off?

It is a fair objection. Every brand here is a GEOforge client doing content work, so own-site citations could be rising for that reason alone. We have three pieces of evidence against it, and one that is weaker than we hoped.

• ChatGPT did not move. Own-site share across the same sites and prompts was 3.2% before and 3.1% after, and stayed between 3.5% and 4.1% in every week from 13 July to 28 September. It rose in 4 of 7 sites and fell in 3.
• The presence drop only shows where lists shrank. In AI Mode, 6 of 7 sites appeared in fewer answers. In AI Overviews, where lists shrank by about a fifth, it was 3 of 7.
• Sites whose lists barely changed barely changed. The two regional sites had short lists to begin with and moved least.
• Week to week, the link is real in some sites and weak in others. When a site’s source list got shorter, its own-site share rose in 4 of 7 sites (correlation of weekly changes below −0.3), and showed no clear relationship in the other 3. In ChatGPT, where lists did not change, the same measure was mixed: 2 negative, 3 positive. With 12 weeks per site, we would not lean on any single correlation.

Is this just one or two brands?

No, though it is a small panel. The seven sites include six brands, and no site accounts for more than 27% of the own-site citations in the after period. Dropping any one site in turn, the average own-site share still rose and the average presence still fell, seven times out of seven. Counting the two regional sites as one brand gives 3.0% to 4.8% for share and 38% to 30% for presence. Dropping Site E, the one with a short history, gives 2.5% to 4.0% and 30% to 24%. Including the few prompts that exist on only one side changes only Site E, and moves its numbers to 5.2% to 7.6% and 55% to 48%, the same direction.

One site is the exception on presence: its share of answers citing the brand rose, from 5% to 15%, from a very low starting point. We include it in every average.

Was it one step, or a drift?

Not one clean step. Sources per answer fell from about 19 to about 10 in the last week of August, then dropped again in mid-September to about 6 and partly rebounded. Own-site share moved with it in direction, but not week by week: after the cut it bounced between 3.4% and 4.4% rather than tracking each wobble in list length. That is why we use a before-and-after comparison on matched prompts as the main result and treat the weekly chart as context.

What should you do with this?

• Track presence and share separately. A rising share of citations looks like good news. In AI Mode it came with the site appearing in fewer answers. If your dashboard shows only share, you can miss losing ground.
• Treat a short source list as a scarcity problem. With about 8 sources per answer instead of 18, each slot matters more. Third-party sites lost the most, so earning citations on other sites has less room to work in AI Mode than it did.
• Check each engine on its own. ChatGPT and AI Mode behaved differently over the same weeks for the same sites and prompts. A blended number hides that.
• Do not read this as a cause. We saw what happened to brand sites when lists got shorter. We did not test why Google changed them, or whether any specific kind of page does better.

For the earlier shift we saw in the same week, see why LinkedIn posts overtook articles in Google’s AI answers.

How did we measure this?

Data. Every AI answer GEOforge collected between 6 July and 4 October 2026 for seven brand sites, restricted to answers that came with a source list, to prompts measured on both sides, and excluding 17–30 August. That is 53,274 AI Mode answers; ChatGPT and AI Overviews were measured the same way. We counted the sources each engine attached, not links inside the answer text.

Own site. A source counts as the brand’s own site if its host is the domain of the brand’s website or any subdomain of it. “Presence” is the share of answers with at least one such source.

Normalization. Shares of citations and of answers rather than counts, so changing volume does not matter. Brand averages are unweighted means of the seven sites; weighting the two regional sites as one brand does not change the result.

Checks. Before writing this we checked: the matched-prompt restriction; the weekly shape; each engine separately; leaving out each site in turn; the final week; and our own pipeline. A Google redirect problem in late August 2026 had left some answers without real source URLs; those answers have since been repaired, and we dropped any that could not be. We also confirmed that nothing in the code that collects Google or ChatGPT answers changed between 10 August and 20 September. Our June data for AI Mode is too thin to give an earlier baseline: only one full week had enough answers in at least four of these sites.

Limitations. Seven sites is a small panel, and every one belongs to a client of GEOforge, so these brands do GEO work and their sites may be better cited than average. Prompts are the ones these brands track, which skew toward their own categories. We cannot see why Google shortened its lists. Only the five sites whose lists roughly halved show the effect cleanly; the two regional sites barely had a cut to react to. Every figure is aggregated and no brand or prompt is identified.

Sources & method. GEOforge production citation database, 6 July to 4 October 2026 (17–30 August excluded): seven brand sites, matched prompts, 53,274 Google AI Mode answers. Related: LinkedIn posts overtook articles in Google’s AI answers and Engagement Is a Threshold, Not a Ladder.