Answer engines changed what they read in August

Something changed in how answer engines assemble recommendations over the last two months, and most brands have no way to see it happening. Scores drift, nobody can say why, and the usual explanations do not fit. We capture every answer and every source behind it, which means when a brand moves we can usually say what moved it.

We had one client where the cause was unusually easy to isolate, so we will start there. In August, if you asked Microsoft Copilot to name the top brands in their product category and state, it named them about half the time. By early September it named them in roughly one answer in twenty.

Nothing about the brand had gone wrong. Independent industry sales data has them up 59.8% year over year over the same period, climbing from fourth to third in their state category, and stocked in over 700 retailers. The business was having a very good quarter while the answer engines were quietly writing it out of the category.

We measure this for a living, so we went looking for the cause. It turned out to be smaller and stranger than we expected.

One page was doing all the work

Copilot builds its answers out of pages it retrieves. In August, for this category, it kept going back to the same one: a retailer's buying guide ranking the best products in that category, which listed our client fourth and quoted their market share.

Copilot cited that page in 22 of our 51 captures in August. In September it cited it zero times.

The page is still live. It was last updated in July, it still lists the brand, and nothing about it changed. Copilot simply stopped reading it.

What it read instead is the interesting part. We opened every replacement source by hand. The answers had moved from market-share coverage to awards coverage. One replacement names no brands at all. One is a dispensary's awards roundup listing 23 winners, without our client. The third names them, but only in its section for a different state.

Two things make this more than a coincidence

A control, by accident

ChatGPT kept citing that same page over the same period, 27 captures in August and 27 in September. And on ChatGPT the brand's mention rate held steady, 63% to 70%.

Same brand, same page, same weeks. One engine kept reading the document and kept recommending the brand. The other stopped reading it and stopped recommending them. That is about as close to a controlled experiment as this field allows.

The mirror image, in another state

The brand did win an industry award in 2025, in a different state, for a different product line. As the engines shifted toward awards-based sourcing, their visibility in that second state went up across all five surfaces we track, in the same month their home-state visibility fell. Perplexity went from 29% to 52%. Gemini from 5% to 14%.

This is part of something larger

August was an unusually busy month for retrieval changes, and the case above was one brand caught in the crossfire of one of them. The wider pattern shows up across our whole book.

The most reported was ChatGPT's collapse in Reddit citations, which several trackers put between 73% and 86%^[1]. Our own data agrees and adds two things to it. First, it was not selective. Restricting to brands with enough volume on both sides of the date to compare each against itself, every single one fell, across cannabis, dental practices and home services alike. The mean went from 24% of answers citing Reddit to 2%.

Second, and less remarked on, Microsoft Copilot has never cited Reddit at all in our data. Not once, in any week since June. Advice to “go win on Reddit” was never advice about Copilot.

OpenAI has said it does not target individual sites^[2], and the most credible account is a change in how ChatGPT searches rather than a decision about Reddit: on 8 August the share of its internal fan-out queries using a site: operator jumped roughly fortyfold in a single day^[3], and the number of searches per answer nearly doubled. It started asking named sites directly instead of casting a wide net.

Only one source mattered

Reddit fell 7.16 percentage points of citation share and it fell for every single brand we could measure, without exception. Everything else people assume sits in this category was already negligible before any of it happened.

Wikipedia was 0.07% of citations before any of this. Quora did not appear once in 251,846 measured answers. Reddit was the user-generated category, and its removal is the whole effect.

How exposed you were depended on your category, not on anything you did

The size of the loss tracked almost perfectly with how much a category gets discussed by consumers. A cannabis brand lost 20.8 points of citation share. A commercial real estate firm lost 0.33. That is a sixty-fold spread, and no brand did anything to earn or avoid it.

What took its place

No single source absorbed it. Three classes did, and the largest share went to a long tail rather than to any one destination.

Directories took 2.11 points and government sources 1.41. The remaining 3.64 points went to trade publications, peer businesses and manufacturers. The most consistent single gainer anywhere in the data was reviews.birdeye.com, which rose for all ten brands where it appears. Everything else was vertical-specific: LoopNet and Crexi for commercial real estate, Angi for home services, Realtor.com for residential, state licensing boards and association finders for regulated trades.

And the engines are diverging on how many brands they name

Separately, the number of distinct brands an answer names has been moving, and not in the same direction everywhere. This has to be measured per brand, comparing each against itself. Category shape varies enormously, a question about dental practices names a different number of competitors than one about live resin, so any figure pooled across a changing mix of brands measures the mix rather than the trend.

Both extremes are unanimous. Every brand we could measure on Google AI Overview saw it name fewer competitors. Every brand we could measure on Copilot saw it name more. Those are not noisy trends with a few outliers, they are the whole cohort moving together in opposite directions.

We have left Perplexity out of this comparison. A collection outage through late August makes its figures for this period unreliable, and we would rather say so than quote a number we do not trust.

What we would actually do about it

The uncomfortable implication of all this is that being genuinely good is necessary and not sufficient. A brand can be growing 60% a year and still disappear from an answer, because the answer is assembled from documents, and the brand is not in the documents currently being read.

Know your anchor sources. One URL was carrying 95% of this brand's Copilot visibility in its most important category, and nobody knew that page existed until it stopped being cited. That is a concentration risk, it is measurable, and almost nobody is measuring it. If a single source is doing most of the work for you, you want to know before it goes quiet, not after.

Treat each engine separately. Reddit is worth something on Perplexity and Google, is worth much less on ChatGPT than it was in July, and has never been worth anything on Copilot. Copilot retrieves through Bing's index, which means a page's standing in Bing is now a separate problem from its standing in Google. A single blended strategy hides all of this.

Get into the coverage the engines are currently reaching for. Right now that leans toward awards lists, recent industry coverage and current market reports, and away from evergreen roundups. That will change again. The durable version of this advice is not “chase awards,” it is “know which documents are answering your category this quarter, and be in them.”

None of that is possible by reading your own score. You have to look at what the engines actually cited, answer by answer, week by week. That is the part we build.