In the single week to 30 August, the number of unique pages AI engines cited fell 47.3% across seven industries, and fell in all seven. Two engines lost more than four fifths of their citations. Three, answering the same questions in the same week, did not move at all.
Research
Two consecutive weekly snapshots, 23 and 30 August 2026. A single week, because that is where the movement is. Every number below is unpacked in the sections that follow.
The dates, the markets and the scale behind every number in this report.
This study compares two consecutive weekly snapshots, 23 and 30 August 2026, across roughly 661,000 measured prompts in the United States and Canada, in seven industries: health insurance, consumer goods, personal care, banking, insurance, veterinary diagnostics and telecommunications.
Two measures run side by side. Unique pages cited counts how many distinct pages an engine drew on. Citations counts every time a page was used, so the same page cited in twenty answers counts twenty times. They answer different questions, and in this week they moved by very different amounts.
Answer counts were identical from one run to the next in every industry, so a change here is a change in how many sources an answer carries rather than a change in how much was asked. The four-week series in the last section runs on the six industries with a complete comparable history.
The count of distinct pages the engines drew on fell 47.3% in seven days, and fell in every industry measured.
Research
All engines pooled, 23 to 30 August. Zero sits at the right edge, so every bar runs leftward. The bottom row is all seven industries pooled together.
323,760 pages that were being cited on 23 August were not being cited on 30 August. The severity varies by less than a factor of two, from 32.4% in insurance to 56.4% in personal care. The direction does not vary at all.
This is the measure that matters most if your strategy is built on being present somewhere in a wide pool of credible pages. The pool halved in a week.
ChatGPT citations fell 81.4% and Google AI Mode fell 83.2%. The second engine had been flat for three weeks, so this is not one engine unwinding a spike.
Research
23 to 30 August, same answers, same week. Telecommunications does not run Google AI Mode, so that industry appears in the ChatGPT column only.
| Industry | ChatGPT | Google AI Mode | Unique pages cited |
|---|---|---|---|
| Banking | −84.7% | −79.8% | −34.8% |
| Insurance | −84.4% | −80.8% | −32.4% |
| Health insurance | −83.8% | −86.3% | −55.5% |
| Veterinary diagnostics | −83.6% | −76.4% | −44.0% |
| Telecommunications | −82.6% | −42.3% | |
| Personal care | −76.3% | −84.2% | −56.4% |
| Consumer goods | −76.2% | −84.4% | −43.2% |
| All industries pooled | −81.4% | −83.2% | −47.3% |
ChatGPT went from 848,332 citations to 157,658. Google AI Mode went from 590,441 to 99,143. Neither is a marginal move and both land in the same narrow band across every industry.
The second engine is the one that settles what this is. Google AI Mode carried a steady 21.4, then 20.7, then 20.1 sources per answer over the three weeks before this one. It had no recent rise to unwind. It fell anyway, and it fell slightly further than the engine that did.
The two measures also separate cleanly. Unique pages fell 47.3% while citations fell more than 80%, which means the engines are drawing on about half as many pages and leaning on each surviving page far less. Both effects compound in the same direction.
Same questions, same reports, same seven days. This is what an engine that did not change looks like, and it is the reason the two that fell can be read as a change in the engines rather than in the market.
Research
All seven industries pooled, 23 to 30 August. The version-pinned engine is a frozen build that cannot respond to a vendor release, so it sets the scale for measurement drift.
Perplexity and Copilot finished within 5% of where they started. The frozen build finished 14.0% higher, which is the honest width of measurement drift in this window. Google AI Overview fell 15.4%, closer to that drift than to the two engines that lost four fifths of their citations.
Both falling engines held their answer counts exactly, and both named more brands than the week before, not fewer. Brand visibility rose from 72.0% to 74.3% in one and from 67.3% to 69.8% in the other. So they are writing the same volume of answers about the same brands, and attaching far fewer sources to them.
That combination is what rules out the ordinary explanations. An engine answering less would show fewer answers. A brand losing ground would show lower visibility. A measurement fault would not spare three engines running through the same reports on the same days.
Citations rose sharply in mid-August. Reading the fall as that rise unwinding gets the level wrong by a wide margin.
Research
Six industries with a complete comparable history, pooled. Each industry's 9 August level is set to 100.
Pooled across those six industries, ChatGPT citations finished 71.4% below their 9 August level, and finished below it in 6 of 6. The fall did not stop at the pre-rise level. It went straight through it.
The same holds for what kind of page gets cited. Independent-publisher share of ChatGPT citations was lower on 30 August than on 9 August in all six industries, from 74.5% down to 68.2% in health insurance and from 59.9% to 41.8% in banking. So the shift toward pages that brands and their competitors publish directly, which we reported in The Direct Line, survived the collapse in volume.
Two things moved in opposite directions this month, and only one of them is worth setting a target against. Volume swung from 100 to 169 to 29 in three weeks with no action from any brand. The mix of who gets cited moved once and stayed.
Three conclusions that follow from the evidence above, and the honest limit of a one-week window.
It moved from 100 to 169 to 29 in three weeks while every brand in the panel did nothing. A metric that swings six-fold on vendor behaviour cannot carry a target. Report share of what is cited, which moved once and held, and keep volume as context.
The pool of pages the engines draw on halved. A plan that depends on being present somewhere across a very wide set of third-party pages now has half as many slots to be present in. Depth on pages the engines still read is worth more than reach across pages they have stopped reading.
Two engines lost more than four fifths of their citations. Three moved less than 5%. Averaging those together produces a figure that describes none of them, and it would have hidden the largest citation movement we have measured.
Two consecutive snapshots establish that this happened in that week, that it is confined to two engines, and that it is larger than the drift of a frozen build by a wide margin. They do not establish where it settles. We have already measured one citation contraction in this data that recovered fully and overshot inside three weeks, so a fall this steep is a level we will treat as provisional until the next snapshots confirm it. The direction and the isolation are what we are confident about now.