Gemini dropped Reddit too. It wasn’t a ChatGPT story after all.

What moved

Compared per engine — chatgpt against August 2026, gemini against July 2026 — because a wave can cover one engine and not the other.

In August, ChatGPT stopped citing Reddit for local-business questions. I said at the time it looked specific to ChatGPT, because a 40-query Gemini spot-check came back near its normal range. That was wrong, and this wave is what corrected it. Gemini did the same thing, roughly five weeks later.

Reddit fell out of Gemini too

reddit.com went from 13.7% of every Gemini citation to 1.7%. In July it was the single most-cited domain in the entire study, ahead of BBB, HomeAdvisor and Angi combined. It’s now eleventh. Both major AI engines dropped their biggest community source inside of six weeks of each other, and nobody announced either one.

The vertical split matters more than the average does. Reddit didn’t get zeroed out anywhere. It got flattened almost everywhere. Across home services it ran 11–17% of citations in July and runs about one point now — HVAC 17.0% to 0.9%, plumber 15.0% to 1.3%, electrician 13.7% to 0.5%, with roofer, house cleaning, pest control and locksmith all the same shape. It holds up best where the question is more subjective, auto repair (32.5% → 6.4%) and dentist (9.3% → 3.8%), and even there it gave up most of what it had. The one category that didn’t move is personal injury lawyers, which barely used Reddit before (0.4%) and barely uses it now (0.5%). Nine of ten categories fell, so this is closer to a blanket de-ranking than I first read it — and the two categories that leaned on it hardest lost the most ground in absolute terms.

The directories took it, in both engines

Same destination as ChatGPT in August. Gemini’s local-service directories went from 10.4% to 19.1% and general directories from 10.5% to 17.7%. Angi nearly quadrupled to 6.4%, HomeAdvisor doubled to 4.3%, Expertise more than doubled to 3.5%. Facebook shows up at 2.1% having been absent in July. On the way out: care.com, zocdoc.com, youtube.com and deltadental.com.

Businesses’ own sites did not absorb it. That share fell from 58.8% to 49.6%. If you want one sentence for both engines: when the forums went, the directories won.

Gemini is citing less, from fewer places

Every query now comes back with 6.8 sources on average, down from 9.7. Across the wave Gemini drew on 2,474 distinct domains, down from 3,611, and named 3,738 distinct businesses, down from 4,410. It’s not just swapping one source for another — it’s drawing from a visibly narrower pool.

Which shows up in the repeatability test

Ask Gemini the identical question five times and the citation sets now overlap more than twice as much as they did in July — 0.177 to 0.383. That’s a big enough jump that I’m comfortable calling it real, and the narrower pool explains it. Fewer places to draw from means repeat queries land on the same sources more often.

Which business it names first went the other way, 7.9% to 5.1%. I’m flagging that one rather than claiming it. A 2.7-point move doesn’t clear the three-point bar this study uses before it calls anything a change, so by my own rules it sits inside the noise. Worth watching next wave. Not worth building on.

One caveat I owe you, because it cuts against my own framing. Google’s local pack — the control, the thing that’s supposed to hold still — also loosened, from 90.2% to 83.3% on the same top result. Finding Eight’s contrast still holds at roughly 16x, but “one surface holds still” is now an overstatement. The control moved too.

Rewording a question costs you nothing

This is the one I keep coming back to. Ask Gemini the identical question five times and the citation sets agree 43.0% of the time. Ask the same underlying question three different ways — “best plumber in Dallas,” “top rated plumber near Dallas,” “who is a good plumber in Dallas” — and they agree 41.2%. Those are the same number.

Which means phrasing isn’t the problem. I built the whole phrasing-volatility test on the assumption that different wording pulls different sources, and it does — but no more than identical wording already does. You could type the exact same query twice and do about as badly as someone who worded it completely differently. It held in July too, and pointed the same direction: identical questions agreed 36.8% against 40.3% for reworded ones, so back then asking the same thing twice was actually worse than rephrasing. Either way, there’s no phrasing you can find that stabilises the answer.

The next-day decay in the drift study didn’t reproduce

The Grounding Drift page has said since July that citation agreement gets worse overnight — 46.3% on same-day repeats, 26.5% the next day. I re-ran that at wave scale, 500 queries a round instead of six. Round pairs 30–40 hours apart agree 42.7%; pairs two to ten hours apart agree 43.0%. The difference is 0.2 points, with a confidence interval of −0.9 to +1.2. There’s nothing there.

July’s version was real inside its own sample — all six queries moved together and it clears a significance test — it just doesn’t generalise past six queries in one city. I’ve retired the decay reading and corrected the page. Two things I got wrong there, for the record: I quoted the two highest of the previous day’s three rounds and left out the lowest (0.332), which made the next-day number look like more of an outlier than it was, and I presented a six-query result with more confidence than it had earned.

What survives is the part that never depended on time. Gemini’s own internal search strings agree with each other only 0.011–0.056 of the time, far less than the domains that come out the other end. The instability starts at the moment the model decides what to type into Google, and it’s steady. There’s no quiet period to wait for.

What didn’t change

Twelve source categories came back inside the noise floor, and so did most of the headline findings. Average rating of an AI-recommended business: 4.75 → 4.76. Cross-phrasing agreement: 40.3% → 41.2%. Gemini-vs-ChatGPT domain overlap: 8.3%, unchanged to the decimal. Same top business named by both engines: 4.2% → 4.9%. Every one of those ran through the same test as the movers and came back flat.

One I’ll report but won’t file in that column: the share of businesses turning up in more than one metro went 2.0% to 1.8%. That stat doesn’t go through the change test, so I can’t tell you whether it moved or just wobbled. Small either way, but I’d rather say which numbers were tested than let them all borrow the same confidence.

That matters as much as the movement. A wave where everything moves is usually a broken measurement, not a finding.

I re-audited my own classifier, and it changed the August numbers

Every domain gets sorted into a source category, and anything the classifier doesn’t recognise falls back to “the business’s own website” — reasonable, since the long tail of local-service citations genuinely is thousands of individual company sites. The risk is that a high-frequency unrecognised domain is almost never one business; it’s an aggregator with a company-sounding name.

ChatGPT’s own-site share had climbed 15.9% → 42.8% → 48.4% across three waves. That’s exactly the drift the fallback is vulnerable to, so I re-audited the highest-frequency domains by visiting each one. Roughly three dozen were misfiled: reviews.birdeye.com, bestprosintown.com, avvo.com, attorneys.superlawyers.com and peers are review platforms and ranked-list directories, not businesses. ChatGPT’s true July own-site share was 2.9%, not 15.9%.

All ChatGPT figures for every wave are restated on that basis, which sharpens rather than softens the story: the share Reddit vacated went to directories and review platforms, and businesses’ own sites were the smallest of the four gainers. Gemini was barely affected — only 1.7 of its 49.6 own-site points came from named domains, and those are genuine franchise sites.

Every aggregation run now prints any unrecognised domain above 0.5% of a wave’s citations, so this can’t drift quietly again. That list is empty for all three waves as of this build. The mechanism is written up under Source classification.

What I’d do about it

If your AI-visibility work was built on community presence, the ground moved under it twice this quarter, in both engines, with no notice either time. What replaced it is unglamorous and buyable: Angi, HomeAdvisor, Expertise, Birdeye, BBB, Thumbtack. Being listed, rated and ranked there is now most of what both engines see.

The honest caveat: I’ve watched this shift twice in six weeks. Anyone selling you a durable AI playbook off a single month of data — me included — is selling something the systems themselves don’t support. That’s why this page exists and why the numbers get re-measured rather than restated.

ChatGPT dropped Reddit from local-business answers. Completely.

On July 30, 2026, reddit.com was 41.7% of every source ChatGPT cited when I asked it for a local business — the single biggest source in the whole ChatGPT dataset. Then in mid-August the reports started: ChatGPT Search had sharply cut back on Reddit, around the same time it began running site:-scoped background searches at scale (roughly August 8). So I re-ran the identical 1,500-query design on August 20.

What moved

reddit.com went from 41.7% to 0.0%. Not a decline. Zero citations out of 3,452. A same-day follow-up pass of 150 queries found the same thing — 0 of 334 — so it isn’t a one-off snapshot. September made it three measurements in a row at zero.

Where the share actually went

Directories and review platforms took nearly all of it. Local-service directories went from 6.7% to 23.0%, general business directories from 39.1% to 51.6%, and review platforms — barely present in July at 1.3% — jumped to 12.0%. Businesses’ own websites gained too, from 2.9% to 9.7%, but that’s the smallest of the four moves, not the story.

Look at the individual sites and it’s clearer still. In September, expertise.com alone is 21.8% of everything ChatGPT cites for these queries, reviews.birdeye.com 10.2%, bestprosintown.com 7.5%. ChatGPT swapped one dominant source for a handful of ranked-list directories and review-hosting platforms. It didn’t redistribute Reddit’s share to the businesses themselves.

Why the July number was so high to begin with

Reddit reliance was never uniform. Dentist queries leaned on it for 72.2% of their citations and auto repair for 60.1%; personal injury lawyer queries only 12.3%. Word-of-mouth questions pull community discussion. Legal questions pull directories. That’s the same pattern Finding Three found in Gemini’s data.

External check

Promptwatch’s general-web tracking saw a smaller, partial drop over the same window: 3.83% (July 18–August 7 average) to 0.52% through mid-August, an 86.4% decline. Local-business queries got hit harder. One plausible reason: they’re easy for a site:-scoped search to route straight to a directory listing, while open-ended, opinion-seeking queries may still surface Reddit threads. Plausible. Not tested — I haven’t run a matched set of open-ended queries.

Is it just ChatGPT?

Yes. A 40-query Gemini sample the same day put reddit.com at 19 of 284 citations (6.7%), in line with Finding Three’s 13.7% baseline once you allow for the small sample. This one belongs to ChatGPT alone.

Is ChatGPT still using Reddit and hiding it?

One theory going around says ChatGPT still fetches Reddit threads internally, then credits vendor pages instead. That evidence comes from intercepting live browser traffic, which I can’t replicate through the API. What I could check: does Reddit show up anywhere in the visible output without a citation link — the answer text, the named brand entities, known Reddit-mirror domains? Across all 1,650 post-drop queries, zero. Not one mention of Reddit or subreddit-style phrasing. For comparison, 899 of the 1,500 July 30 queries (60%) mentioned Reddit somewhere in the answer text back when it was being cited. I can’t rule out Reddit shaping the model’s reasoning invisibly — that would take network-level access this study doesn’t have — but if it’s happening, it leaves no trace anywhere a user or an API caller can see.

What it means for a local business

If your ChatGPT visibility strategy was “get talked about on Reddit,” it stopped working overnight, and you didn’t get a memo. What replaced it isn’t your website — it’s a short list of ranked-list directories and review platforms. Being listed, rated and ranked on expertise.com, birdeye, Angi and their peers is now most of what ChatGPT sees when it answers these questions. That’s a different job than building a good website, and worth saying plainly: for ChatGPT specifically, the directories won.