Google AI Mode cited 67% fewer sources, and 11 brands lost a third of their visibility in the same week
None of those brands did anything. What changed was the search engine.
Between 24 and 31 August 2026, Google AI Mode went from citing 12.22 sources per answer to 4.08 across the same 990 questions, and its answers got 32% shorter. Share of Answer for the 11 brands measured fell from 23.4% to 15.9% in that same week.
- Sample
- 4,852 Google AI Mode answers · 990 questions per week · 11 brands across 10 sectors
- Period
- 3–31 August 2026 (weeks 32 to 36)
- Method
- Closed panel: only brands measured in all five weeks. Same questions, same transport, one repetition per question. The method stamp is identical across all five weeks.
- Language
- Spanish (Spain)
What was measured
The same 990 questions, for the same 11 brands, five weeks in a row. It is a closed panel: if a brand is not present in all five weeks, it does not count. Without that filter, a brand appearing or dropping out moves the average and you can no longer tell an engine change from a sample change.
| Week | Answers | Citations / answer | Distinct domains | Length (chars) | Sector brands named | Share of Answer | Share of Mentions |
|---|---|---|---|---|---|---|---|
| 2026-W32 | 988 | 13.07 | 11.06 | 4,310 | 1.06 | 22.3 % | 17.3 % |
| 2026-W33 | 894 | 12.82 | 10.82 | 4,481 | 1.04 | 22.8 % | 18.0 % |
| 2026-W34 | 990 | 12.22 | 10.43 | 4,323 | 1.00 | 23.4 % | 19.0 % |
| 2026-W35 | 990 | 9.94 | 8.33 | 3,854 | 0.94 | 23.5 % | 19.9 % |
| 2026-W36 | 990 | 4.08 | 3.46 | 2,918 | 0.61 | 15.9 % | 20.5 % |
Nothing changed on our side between week 34 and week 36: same questions, same transport, one repetition per question. The method stamp stored with every answer is identical across all five weeks, so the series is homogeneous and comparable with itself.
The answer shrank, and the shelf shrank with it
It is not just that there were fewer citations: there was less of everything.
The four middle columns move together. The average answer went from 4,323 to 2,918 characters (−32%), distinct domains cited from 10.43 to 3.46 (−67%), and sector brands named in each answer from 1.00 to 0.61 (−39%). A shorter answer names fewer companies. All of them.
The number that rules out the comfortable explanation
The easy reading of a Share of Answer drop is that the brand lost ground to its competitors. Here the opposite happened: Share of Mentions — the brand's share of all mentions in its niche — rose every single week, including the crash week (19.0% → 19.9% → 20.5%).
The 11 brands lost mentions. Their competitors lost more. Nobody lost ground: the shelf got smaller.
Why this is not our instrument failing
That is the first hypothesis to rule out, and three things rule it out:
- The method stamp is the same across all five weeks. We changed neither transport, nor model, nor number of repetitions.
- The data provider behind the transport confirmed it in writing on their own sample of roughly 2,378 paired daily prompts: the drop starts on 22 August and their citation markers match the extracted sources one for one (1.87 versus 1.86). If links were being lost during extraction, there would be more markers than sources.
- The drop starts four days before Google changed the format of its outbound links on 26 August. That change explains the low point on the 27th; it does not explain the step already underneath it.
Our panel is independent of theirs — different questions, different language, different sectors — and reproduces the same pattern, answer shortening included. Two samples that never talk to each other finding the same thing is as close to a replication as you get here.
Why this matters more than the number itself
On 24 August 2026, a public AI visibility index showed, across a thousand measured brands, one rising and 987 falling. None of those 987 brands had done anything wrong.
That is the methodological failure this finding illustrates: a tool that reports Share of Answer without keeping a record of what the engine is doing bills a search engine change to the brand. And the mistake is expensive: a client watching their chart fall fires their agency, rewrites their content strategy or cancels the service, over a movement that was never theirs.
That is why we keep a dated catalogue of engine changes, each with its source, separate from the metrics. When a drop coincides with a declared event, the report says so instead of letting the chart talk on its own.
What this finding does NOT prove
- It does not say Google will keep behaving this way. It is a measured step change, not a projected trend.
- It does not measure the effect on traffic. Fewer visible citations may mean fewer clicks, or none: that is measured in analytics, not here.
- The panel is Spanish and covers 10 specific sectors. It cannot be extrapolated to other languages or to markets with a different competitive set.
- It does not identify the cause inside Google. We know what changed in the output, not what decision drove it.
- Eleven brands is a small sample of brands, even at 4,852 answers. What is robust is the pattern, not the exact magnitude per sector.
Frequently asked questions
Will Google AI Mode go back to citing as it did before?
Nobody knows, and anyone who says otherwise does not know either. The provider who confirmed the drop on their own sample describes it as the new normal behaviour, not as an incident. What can be stated is that as of 1 September 2026 there was no recovery.
Does it mean my site lost ground in AI search?
Not necessarily. In this panel, brands lost absolute mentions while gaining share against their competitors. To know which of the two happened to you, look at Share of Mentions, not only Share of Answer.
Does it affect ChatGPT and Perplexity too?
This finding measures Google AI Mode only. Each engine decides on its own how many sources to cite, and in our own corpus two engines answering the same question share between 4% and 13% of their sources.
How can I check whether this happened to my brand?
Compare the same week before and after across the same questions, and check whether the number of sources cited per answer fell at the same time as your presence. If both fell, the suspect is the engine; if only yours fell, the suspect is you.