Ask ChatGPT “How do I file for divorce?” and you’ll see an ad about one time in five. Ask “I need a lawyer, who should I call?” — same broad topic, same platform, same week — and you’ll see one three times in five. That 42-point gap, on the identical nominal subject, turned out to be the single largest and cleanest effect in a study we didn’t design to find it.
We built this study to answer a narrower question, prompted by an incidental finding in our companion study on ChatGPT’s local citations: while counting citations there, we noticed 59.6% of local-business queries had a ChatGPT ad attached, far more than published industry baselines suggested. So we asked: does a “best plumbers near me” query really draw more ads than an ordinary commercial query like “best high-yield savings account”? We ran 1,200 local-intent prompts across 10 verticals and 6 US markets against a same-window, same-method 200-prompt control group of general queries with no local intent at all — and once that comparison raised more questions than it answered, we ran three smaller extensions to chase them down. One of those extensions — testing whether an actual informational question behaves differently from a business-lookup request, even on the same topic — is what produced the 42-point gap above.
The original question got a real answer too, and it’s worth reading in full below, including where our first pass at it got the size of the effect wrong.
Findings at a glance
Finding 1 — query type, not topic, is the strongest lever we found
Before the query-type test, we’d already checked whether phrasing matters within a fixed frame: four ways of asking for the same local-business recommendation — “best plumbers in Chicago,” “who should I call,” “what is a plumber?”, “plumbers near me.” All four landed in a narrow 54-62% band with heavily overlapping confidence intervals. Rewording a local-business query barely moves ad presence.
But every one of those four was still, at bottom, a request for a business recommendation. We ran a sharper test: eight genuinely informational legal questions with no city and no business-lookup framing at all — “What is a power of attorney?”, “How do I file for divorce?” — against the same commercial-referral “I need a lawyer, who should I call” queries from the main study, same broad topic either way.
A 42-point gap, non-overlapping CIs, on the identical nominal topic. This number moved with more data, and it’s worth being honest about that: an initial 40-prompt pass found zero attorney-referral ads on informational queries and read as a clean category exclusion. Tripling the sample to 120 showed the real picture is softer — attorney ads do appear on informational queries, just roughly half as often as self-service legal products (PublicRecords.us, LegalZoom), and the attorneys who do show up look topically matched to the specific question (divorce, custody) rather than broadly bidding on “lawyer.” The corrected finding isn’t “attorneys are locked out of informational queries” — it’s “the rate is much lower, and who’s buying shifts substantially.”
Finding 2 — local-intent queries really do draw more ads than general-commercial ones, but our first pass got the size of it wrong
This was the study’s original question, and the incidental finding that started it: our companion citation study noticed 59.6% of its local-business queries carried a ChatGPT ad, while counting citations for an entirely different purpose — far more than two published industry studies suggested (SE Ranking ~26%, a single-day snapshot from 23 July 2026; Evertune ~13%, collected 21 May–8 June 2026). Reading both studies’ methodology pages directly surfaced two problems with using them as the comparison: neither discloses what ChatGPT account tier or login state they scraped under, and both predate this study’s collection window by weeks to months, during a period ad density reportedly grew fast. Comparing against either risked conflating “local intent draws more ads” with “ads got more common since spring.” The fix: collect a same-window, same-method general-commercial control ourselves, so the real comparison happens entirely within one study’s own contemporaneous data.
Ratio: 1.44x, with confidence intervals that no longer overlap — a real, statistically meaningful gap.
It’s worth being honest about how we got here. Our first pass collected only 100 control-group runs — enough, we thought, to establish a same-window reference point. It wasn’t: that sample put general-commercial ad presence at 56.00%, a confidence interval wide enough ([46.23%, 65.33%]) to overlap the local-intent group almost entirely, and we initially wrote this finding up as disconfirmed — “no meaningful gap.” Doubling the control group to 200 runs tightened the interval and moved the point estimate to 40.50%, a 15.5-point swing that reversed the conclusion. The two-group design was still the right call — it’s what made this catch possible. But the first specific conclusion we drew from it was wrong, and only checking at higher power caught it.
That correction also reframes the gap against the published baselines: the apparent 2-4x difference between the companion study’s incidental 59.6% and SE Ranking’s/Evertune’s 13-26% turns out to be partly a real local-intent effect and partly a time effect — ad density does appear to have grown ChatGPT-wide since spring/summer 2026, and local-intent queries also do show meaningfully more ads than general-commercial ones in the same window. Both are true; neither alone explains the full gap. And set against Finding 1 above, a 1.44x local-vs-general gap is real but modest next to the 3x gap between an informational and a commercial-referral framing of the same topic — location matters, but query type appears to matter more.
Finding 3 — the category gradient is real, but it doesn’t split the way paid-search intuition predicts
Dentists, lawyers, and plumbers land high, matching the “high paid-search-intent” intuition. But so do hotels and restaurants — categories with no particular reputation for aggressive paid-search bidding — and florists lands mid-pack rather than low. The category story is real but messier than “commercial-intent categories beat discovery categories.”
Auto repair is the sharp exception, and we checked why. At 18.33%, it’s less than half of the next-lowest category and roughly a quarter of the top. This isn’t a fluke of one small sample — our companion citation study found the identical pattern independently, with auto repair dead last among its 12 verticals too (30.42%, a full 10 points below the next-lowest). Two studies, different designs, different windows, same result.
We checked whether this is a rendering problem or a supply problem. It’s supply: auto repair’s ad fill rate (rendered ÷ served) is unremarkable at 55.8% — several other verticals fill worse. What’s different is that auto repair queries were served only 0.36 ad slots per run, roughly a third of the next-lowest category and a quarter of the top categories. OpenAI’s ad system simply isn’t finding or serving nearly as many ads for this category — a demand-side gap, not a display-side one. And when ads do appear, a few are loose fills rather than targeted matches (an appliance-repair company on an auto-repair query, a car-sharing marketplace, a solar retailer) — consistent with what SE Ranking separately found about relevance dropping in thin-advertiser-coverage categories. For a business in this category, the honest read is either “thin competition, real opportunity” or “your peers haven’t caught on to ChatGPT ads yet” — not a measurement artifact.
Finding 4 — small towns show a real, growing effect; regional-metro tiers don’t
This study’s and its companion study’s original market sets stratified by tier — dense urban (New York, San Francisco), mid-size metro (Nashville, Raleigh), and “small metro” (Boise, Fargo) — and the ad-presence gradient across those tiers was essentially flat (56.5-60.75%, heavily overlapping CIs). But none of those “small” markets are actually small; they’re all real regional metro areas of well over 100,000 people.
We tested four genuinely small towns instead — Hood River, OR (~7,800), Sheridan, WY (~18,000), Marquette, MI (~19,000), Dodge City, KS (~27,000) — deliberately avoiding famous resort towns to keep this a fair small-town test.
A real, clean drop that got more pronounced, not less, with more data — an initial 120-run pass found 36.67%; doubling the sample moved it to 30.83%, with non-overlapping CIs against normal markets either way. The mechanism is the same one behind the auto repair outlier: lower ad supply (0.49 slots served per run vs. 0.80 for normal markets), not a worse fill rate. And the advertisers shift too — national wide-net brands (Expedia, Hertz, Airbnb, Wayfair) dominate small-town results, while individual named local businesses, common in normal-market results, are far rarer. National brands can profitably run broad geo-targeting that happens to catch a small town; a business that would need to specifically target 18,000 people mostly isn’t bidding there.
Per-town results remain too imprecise to break out individually even at n=60/town — Hood River reads highest (38.3%) but its confidence interval still overlaps the other three towns. Only the aggregate small-town effect is solid.
Finding 5 — how you phrase the question moves ad presence almost as much as what you’re asking, and we found out why
Everything above tested query type and query category. We hadn’t tested phrasing style — the difference between a short, templated lookup (“What are the best plumbers in Chicago, IL?”) and a real, situational request the way someone might actually type it: “I have a running faucet in my kitchen that won’t shut off and water is pooling on the floor. Where can I find a plumber in Chicago who can come out today?” Same vertical, same city, same underlying need. We wrote ten of these — one per vertical, each grounded in a specific, plausible scenario — and ran them against the closest templated equivalent from the main study.
Pooled: 23.81% for naturalistic phrasing vs. 58.0% for templated — roughly a 2.4x gap. Ten scenarios per vertical is a small sample and the individual confidence intervals are wide, but the direction held in all ten verticals with no exceptions — the odds of that happening by chance are under a tenth of a percent. Three verticals (lawyers, dentists, furniture stores) drew zero ads at all across ten-plus real calls each, against a 40-77% templated baseline in the same vertical.
We didn’t leave this as a curiosity — we checked the raw responses to find out why. Two things stood out. First, it’s a supply problem again, the same shape as auto repair and small towns above: naturalistic queries were served 0.35 ad slots per response versus 0.90 for templated ones, while the fill rate (rendered ÷ served) was nearly identical between the two — 67.6% versus 66.4%. Whatever’s happening isn’t a rendering issue.
Second, and this is the actual mechanism: just over half of the naturalistic responses (52%) triggered a different kind of ChatGPT response entirely — a plain, structured list of business names and addresses, with no ad anywhere in it, instead of the usual prose-plus-ad shape. Every single one of those responses had zero ads. The other 48%, where that list didn’t appear, scored 50% ad presence — right in line with the templated baseline. We checked whether this was specific to naturalistic prompts and found the identical pattern in the small-town data from Finding 4: 45% of small-town responses trigger the same ad-free business-list format, also at exactly 0% ad presence. That response type never once appeared in the main study’s 1,520 templated, general-commercial, or informational-query runs.
Who’s actually buying these placements
Local-intent group’s top advertisers: Expedia (69), Thumbtack (53), UrbanStems (51), Priceline (32), GoTickets (24), DoorDash (22), then several individually named local dental and legal practices, Wayfair, trivago, Grubhub.
General-commercial control’s top advertisers: BestMoney, Insurify, Rocket Mortgage, Fetch Pet Insurance, SoFi, LegalZoom, Liberty Home Guard — all national finance/insurance/legal-tech platforms, no individual local businesses at all.
There’s a real qualitative difference here — the local-intent group does include individual named businesses (a specific oral-surgery practice, a specific law firm) bidding directly, something the control group shows zero of. But the local-intent group’s list is dominated by delivery apps and travel OTAs, categories with nothing to do with the “local service lead-gen” story — outweighing Thumbtack and the individual local businesses combined. The honest read: local intent draws a different mix of advertisers, not a more local one on the whole.
Ad quality — mostly on-topic, rarely stacked, rarely defensive
Three more checks against the already-collected corpus, no new data needed.
Relevance. A naive keyword-match heuristic flagged 14.3% of ads as mismatched to their vertical — a number not worth trusting as reported. Hand-reviewing every flagged case showed most were false positives from brand-name blindness (Marriott, Pottery Barn, Philips Sonicare, and several named law firms all flagged simply because the heuristic didn’t recognize the brand) or legitimate travel/lifestyle cross-selling (Expedia or AllTrails on a museum query makes sense as “things to do while you’re in town”). The real mismatch rate, after hand review, is roughly 1-1.5% (7-10 of 714 rendered ads) — a pet clinic ad on dentist queries, an online K-12 school ad on a museum query, a couple of clearly off-topic retail ads on bookstore and auto-repair queries.
Crowding. Two or more ads in the same response is rare — 0-5% across most verticals, highest for lawyers at 5.0%, consistent with personal-injury law’s famously aggressive ad spend carrying over from traditional search.
Defensive bidding. We checked whether any advertiser’s name also appeared in the same response’s organic recommendation — a business paying to defend a spot it already earned for free. Only 5 of 714 rendered ads (0.7%) show this, and every instance is a large national brand (UrbanStems, Expedia), never a small local business defending its own name.
Predictions, scored
We wrote down what we expected before collecting anything, including our confidence in each guess.
The real lesson: check your control group at the sample size the claim needs
Two of this study’s headline numbers — the query-type gap and the local-vs-general gap — flipped or substantially moved between an initial small-sample pass and a follow-up at 2-3x the sample size. Neither initial read was fabricated or careless; both were exactly what you’d expect from real, honestly-collected data at a sample size that turned out to be too small for the precision the conclusion needed. The fix wasn’t a better statistical test — it was more data, aimed specifically at the arms that were sized smallest by design.
If there’s a transferable takeaway, it’s the same one our companion citation study landed on from a different angle: treat any single-pass finding as provisional until it’s been checked at the power the claim actually requires, especially when a confidence interval is wide enough to almost, but not quite, touch the thing you’re comparing it to.
Methodology — how this data was collected and checked
Every number in this piece comes from directly querying ChatGPT and recording exactly what it returned, including any ads — not from a survey, a panel, or a third-party ad-tracking dashboard.
An “ad” only counts if ChatGPT actually rendered it as part of the response — a slot the ad system considered but never displayed doesn’t count. That distinction matters because it’s the difference between measuring what ChatGPT’s ad system is willing to serve and measuring what a real user would actually see.
Before trusting the topline numbers, we checked whether this collection method reflects what a real logged-in user experiences. Three independent checks agreed: an automated sample batch confirmed the ads look like genuine tracked placements, not test data; a manual, logged-in browser check on matching queries turned up real ads explicitly labeled “Sponsored”; and the data provider’s own support team confirmed directly, in writing, exactly how collection works — Free ChatGPT tier, incognito mode, a brand-new session for every single query. The one nuance worth flagging: a fresh incognito session doesn’t carry the ad frequency-capping or personalization history a real ongoing account builds up over time, so any one specific logged-in user’s day-to-day experience may vary somewhat from the aggregate rates reported here.
See the stats/study in context: Steady Demand Research Index
