The Business Website Isn’t Dead

• 50% of citations point to the business’s own site
• 10,229 citations analyzed
• 50 of 50 metros covered

If you assumed AI answer engines mostly relay directory listings and skip individual company sites, guess again.

Share of all 10,229 citations, by source category.

The margin is narrowing, and quickly. In July the own-site share led that combined group by about 22 points; it now leads by roughly 7, because the directories absorbed what Reddit lost. The ranking is the same and the trend is not, so read this as a lead that is shrinking rather than a settled fact.

It still matters, because the fundamentals haven’t changed as much as the “AI killed SEO” crowd wants you to believe. Gemini’s grounding tool is actively retrieving and citing individual company pages, not just relaying whatever already ranks on a handful of aggregator sites. A plumber running a thin, generic website is still at a real disadvantage — on-site content, clear service pages, and a crawlable structure are doing real work here, not getting bypassed.

→ See the full category breakdown on the Dashboard · How this classification was audited

The Quality Bar Isn’t AI’s Invention

• 4.76 average star rating, AI-recommended
• 4.82 average star rating, plain Google search baseline
• 91% rated 4.5+ stars

I took the actual businesses Gemini named in its answers and looked each one up on Google to see its real star rating and review count. On its own, the number looks like proof AI is hand-picking excellence: 98% of recommended businesses sit at 4.0 stars or better, 91% at 4.5+.

But a number alone doesn’t tell you whether that’s AI being picky, or just what “plumbers in Phoenix, AZ” already looks like on Google. So I ran the control group: a plain, unranked category search on Google Places for the same metros and verticals (9,321 businesses), independent of anything Gemini ever cited.

AI-recommended businesses (n=3,567) vs. the plain-search baseline (n=9,321).

The baseline is higher-rated, not lower — 4.82 average stars and 99.1% at 4.0+, against 4.76 and 98% for what Gemini actually cited. Bootstrap resampling puts that gap at 0.05-0.07 stars in the baseline’s favor, and the interval doesn’t cross zero — that’s a real difference, not noise. In other words: AI isn’t applying a stricter quality filter than Google’s own local search already does. The “cream of the crop” look of Finding Two’s percentages was inherited from the search layer underneath. It’s not some special extra layer of AI curation.

Separately, within the AI-recommended set: being recommended more often — showing up across more metros and phrasings — doesn’t correlate with a higher rating (correlation of -0.049, essentially zero), and 17% of recommended businesses have fewer than 50 reviews. So it isn’t a pure popularity contest either. A smaller company with a genuinely strong, if less voluminous, reputation can still get cited.

→ The control-group methodology · How ratings were verified

Reddit Beat the Entire Directory Industry. Then It Was Gone.

This finding reversed, and I’m leaving both halves of it here because the reversal is the useful part.

In July, reddit.com outcited the whole local-service-directory category — Angi, Thumbtack, HomeAdvisor and their peers, combined. 13.7% against 10.4%. I checked whether that was a quirk of the sample and it wasn’t: resampled, the gap ran 2.3–4.4 points in Reddit’s favour and never crossed zero. One forum, no listings, no local-service product, beating the entire directory industry at its own job.

Five weeks later Gemini had dropped it. Reddit is now at 1.7% and the same directory category is at 19.1%. Same test, same resampling, opposite answer: the gap now runs -18.2–-16.4 points, and it doesn’t cross zero either. ChatGPT did the same thing about five weeks earlier. Neither engine announced it.

Both readings were statistically solid. Bootstrap resampling now puts Reddit’s share at 1.4–2.0% and the directory category at 18.2–19.9%, with no overlap. That’s the part worth sitting with: a finding can be real, correctly measured, properly tested — and still be wrong six weeks later, because the thing it measured changed. Nothing about the July analysis was sloppy. The engine moved.

→ See the full most-cited-sources ranking on the Dashboard

There’s No Universal Playbook

The source mix isn’t consistent across verticals. It swings hard depending on what’s being searched for:

A generic “get on the big directories” strategy misses this entirely. Reddit shows up far less often in the personal-injury-lawyer citation mix than it does for home-service verticals — in this dataset, it just isn’t where AI grounding goes looking for legal-services answers. A dentist ignoring Zocdoc and Healthgrades, by contrast, is skipping the two most-cited sources in their entire category. The takeaway isn’t “never touch Reddit if you’re a law firm.” It’s that the source mix is vertical-specific enough that a one-size-fits-all checklist will misallocate effort in both directions.

→ Explore any vertical yourself on the Dashboard

National Platforms (and Brands) Are the Floor

Angi, HomeAdvisor, BBB, ConsumerAffairs, and Reddit itself show up in all 50 metros I’ve collected so far — they’re the baseline every local business competes against nationally. A second tier shows real variation: Thumbtack and Expertise.com appear in most but not all metros, while DiamondCertified and Best Pick Reports show up in a clear minority — present in some regions and absent from others.

The same pattern holds for the businesses themselves. Only 1.8% of the 3,738 unique businesses Gemini named got recommended in more than one metro — and the ones that did (Terminix, Orkin) are national pest-control and service franchises, not independent local companies.

→ Explore any metro yourself on the Dashboard

Grounding Drift

• 41% average agreement across 3 phrasings
• HVAC Contractor most phrasing-sensitive vertical
• Pest Control most stable vertical

I ran three different phrasings of the same underlying question for every metro and vertical (“best X in Y,” “top rated X near Y,” “who is a good X in Y”) and measured how much the cited sources actually overlapped.

Bar length is the average cross-phrasing agreement for that vertical; the thin mark is its 95% confidence interval.

On average, the same underlying question worded three different ways agreed on cited sources only 41% of the time. HVAC Contractor was the most phrasing-sensitive vertical I measured; Pest Control was the most consistent. That specific gap tests out as real (p < 0.0005 on a 2,000-shuffle permutation test) — though worth flagging honestly: that’s the two most extreme of 10 verticals, picked after seeing the data, so the true confidence is more modest than the raw p-value implies. Treat the vertical ordering as a genuine pattern, not a precise ranking.

A follow-up experiment isolates why. I picked six fixed queries and called each one three times in a row with zero wording change at all — the exact same text, back to back.

Even with no wording change whatsoever, the model’s own underlying search terms barely overlapped call to call (as little as 0-7% shared search strings), and the resulting cited domains disagreed almost as much as they did across genuinely different phrasings. I then repeated the same six queries again roughly 3.5 hours later, and again the next day, to check whether real time elapsed makes it worse. It didn’t compound the way you’d expect: the cross-day agreement landed close to the same-day range, not clearly below it (full numbers, including one honest surprise about how noisy the rate itself is day to day, in the deep dive).

So the mechanism isn’t primarily about phrasing sensitivity, and it isn’t the live web changing over the course of a day. It’s that Gemini’s grounding call makes a fresh, stochastic decision about how to search every single time it’s asked — independent of wording, and independent of when you ask.

This finding gets a full dedicated deep dive — the exact repeat-test numbers, why the instability starts in the model’s own search-query generation rather than the live web, independent research from outside this project (including a related, separately-named pattern documented by SISTRIX) that found the same underlying instability using entirely different methods, and a direct control test showing this is specific to AI-generated answers: run the identical repeated-query design against Google’s own local pack (not an AI answer at all), and it returns the same top listing 83.3% of the time, versus 5.1% for Gemini’s generated recommendation on the same queries.

→ How phrasing volatility was measured · Confidence intervals and significance testing

The Long Tail Still Belongs to Individual Businesses

Of 2,474 unique domains cited across the dataset, and 3,738 distinct businesses Gemini actually named, the large majority are individual companies appearing once or twice each. This isn’t a small set of gatekeeping platforms controlling access.

Combined with Finding One, this means citation share in AI answers isn’t consolidated the way, say, Google’s first page of organic results has become. A well-built individual business site can still earn a direct citation on its own, without ever passing through a directory.

→ Browse every individual business and domain in Raw Data

This Is an AI Problem, Not a Search Problem

There’s an obvious skeptical response to Grounding Drift: maybe local search results are just noisy in general, and the same-question-different-answer problem isn’t special to AI at all. I built a direct control test for exactly that: the same 500 (metro, vertical) combinations as this study, repeated over multiple timed rounds, run in parallel against Google’s own local pack — the classic map-pack business listings, not an AI answer — instead of Gemini.

4,996 pairwise round comparisons (Gemini) vs. 4,940 (Google local pack), all 5 rounds.

Run the identical repeated-query design against Google’s own local pack, and it returns the same top listing 83.3% of the time — versus 5.1% for Gemini’s generated recommendation on the same queries. That is not a subtle gap. A conventional ranking system, fed the same local-business data Gemini draws on, stays almost entirely consistent. The instability is coming from how an AI answer engine generates its response, not from the underlying business data or search signals both systems share.

→ The full control-test breakdown · Methodology

Ask a Different AI, Get a Different Set of Sources

• 8.3% average citation-domain overlap between Gemini and ChatGPT on the identical query
• 4.9% of the time, the two engines name the same top business
• 1,500 identical queries run through both engines for a direct, apples-to-apples comparison

Finding Eight showed the instability is specific to AI-generated answers, not local search in general — but that leaves an obvious follow-up: is it specific to Gemini? I ran the exact same 1,500-query design — same metros, same verticals, same literal query text — through ChatGPT’s own search mode, and compared what each engine cites for the identical question.

The two engines don’t just drift independently. They read different parts of the web. Gemini’s single biggest citation category is still the business’s own website (49.6%, down from 58.8% in July). ChatGPT barely goes there at all (10.5%) and leans on general business directories instead (46.0%, versus 17.7% for Gemini).

What changed since July is that community forums stopped mattering to either one. ChatGPT was at 41.7% in July and is at 0.1% now. Gemini went 14.4% to 3.9% five weeks later. Both dropped their largest community source inside six weeks of each other and neither announced it — the story is on What Changed.

What didn’t change is that they still disagree. Ask the identical question of both engines and on average only 8.3% of the actual domains cited are the same one. The top-recommended business matches just 4.9% of the time. They stopped citing forums together and still can’t agree on an answer.

→ The full Gemini-vs-ChatGPT breakdown · Browse every matched query side by side · Methodology

Other Numbers Worth Knowing

Not every number in this dataset earns its own Finding, but plenty are worth having on the record, especially the ones that don’t show up anywhere else. A few of those, pulled straight from the same data behind Findings 1–9:

• 4.9★ median AI-recommended rating vs. 4.9★ baseline
• 66 businesses (of 3,738) appeared in more than one metro
• 500 metro/vertical groups tested for phrasing sensitivity

The rating gap holds on the median too, not just the average. Finding Two showed the plain Google baseline out-rates what Gemini actually cites on average (4.82 vs. 4.76 stars). The median tells the same story: 4.9 stars baseline vs. 4.9 for AI-recommended. But review counts split in an odd way. AI-recommended businesses average more reviews than the baseline (860 vs. 850), while the typical (median) AI-recommended business has fewer than the typical baseline business (307 vs. 311). That crossover points to a heavier tail of high-review outliers pulling the AI-recommended average up, while the ordinary case skews smaller. Separately, 36% of AI-recommended businesses have 500+ reviews — a real number worth having, but not one that changes Finding Two’s conclusion either way.

Grounding Drift, with real examples. Finding Six’s 41% overall phrasing agreement is an average across 500 tested metro/vertical combinations. Here’s what the two ends actually look like. Three different phrasings of the same Roofer search in Louisville, KY came back with only 3% word overlap in the citation set — close to total disagreement. Three phrasings of a Pest Control search in Baltimore, MD, by contrast, agreed 89% of the time — close to perfect consistency, for the identical underlying question.

Gemini and ChatGPT don’t just disagree overall. The size of the disagreement swings by category and by vertical. Finding Nine’s engine-level citation-category split (business site vs. social/community) isn’t the only gap: Gemini cites local-service directories and marketplaces 19.1% of the time, versus just 27.2% for ChatGPT — a third category the two engines weight very differently. And the overall 8.3% domain-overlap figure hides a wide range by vertical: Dentist bottoms out at 3.9%, while Electrician is the closest the two engines get, at 13.7% (still low, just the least-low of the ten). Browse the full category breakdown and every matched query yourself.

What To Do With This

• Don’t abandon on-site fundamentals. Clear service pages, real content, and a crawlable site structure are what gets cited most often when Gemini backs up its recommendations. Investing there keeps you in control of what AI-cited information about your business actually says.
• Don’t assume chasing reviews past a baseline buys AI visibility. Gemini didn’t select a higher-rated group of businesses than a plain Google search already surfaces, and within the AI-recommended set, showing up more often doesn’t correlate with a higher rating. Neither result proves a specific star threshold causes citation — I didn’t hold market, vertical, or prominence constant to isolate rating’s effect — but nothing here suggests review volume, past whatever already gets you into local search results, is buying extra AI visibility.
• Don’t build on community presence right now. An earlier version of this list told you to treat Reddit as a citation channel in its own right. Both engines have since dropped it — Reddit is 1.7% of Gemini’s citations and effectively zero in ChatGPT. Community discussion may still be worth your time for reasons that have nothing to do with AI. It is not currently buying you citations.
• The directories took that share, so claim them. Angi, HomeAdvisor, Expertise, Birdeye, BBB and Thumbtack absorbed almost all of what Reddit lost, in both engines. Unglamorous, and mostly buyable, which is the point — it’s the part of this you can actually act on.
• Build a vertical-specific directory list, not a generic one. Check what actually gets cited in your category before assuming Angi and BBB are enough.
• Check regional and niche directories relevant to your metro — not just the national names everyone already knows to claim.
• Expect Grounding Drift and check more than once. A single “am I cited” test — even repeated word for word — can mislead you either way.
• Re-check periodically, and mean it. This isn’t a fixed rule set. In six weeks both engines rewrote what they cite, without notice, and one finding on this page reversed completely. Anyone selling you a durable AI playbook off a single month of data — me included — is selling something the systems themselves don’t support. Every wave gets re-measured on What Changed.