What actually predicts a ChatGPT ad?
Whether you are asking for information—or for a business recommendation
In our late-August test, “How do I file for divorce?” produced an ad about one time in five. “I need a lawyer, who should I call?” produced one about three times in five.
The topic is the same. The type of request is different. That produced the largest effect in this study: 20.0% versus 61.8% ad presence, a gap of about 42 percentage points.
We originally set out to answer a narrower question. A companion citation study found ads on 59.6% of local-business queries, well above published industry baselines. We wanted to know whether local intent explained the difference.
We tested 1,200 local-intent prompts across 10 verticals and six US markets, plus a same-window control group of 200 general-commercial prompts with no local intent. Three smaller extensions followed when the first comparison raised new questions.
The original question produced a useful result: local-intent queries showed ads 1.44 times as often as the general-commercial control. But the larger observed difference came from query type. Among the legal prompts we tested, recommendation requests showed far more ads than informational questions on the same topic.
By the numbers
| Result | What it measures |
|---|---|
| 20.0% vs. 61.8% | Informational legal questions versus commercial lawyer-referral requests on the same broad topic |
| 58.25% vs. 40.50% | Local-intent queries versus the same-window general-commercial control—a 1.44× ratio |
| 18.33% | Auto repair ad presence, less than half of every other tested vertical |
| 30.83% vs. 54.33% | Small-town versus normal-market ad presence using the same query template |
| 23.81% vs. 58.0% | Naturalistic versus templated phrasing, about a 2.4× difference |
| About 1–1.5% | Estimated relevance-mismatch rate after hand review |
Eight findings
- Query type beat topic. Informational legal questions produced ads on 20.0% of runs; commercial lawyer-referral requests produced them on 61.8%.
- Local intent produced more ads. The local-intent group reached 58.25%, compared with 40.50% for a same-window general-commercial control.
- Category mattered, but not in a simple paid-search order. Dentists led at 75.83%; auto repair was last at 18.33%. Hotels and restaurants also ranked high.
- Genuinely small towns produced fewer ads. Ad presence was 30.83%, compared with 54.33% in normal markets using the same template.
- Naturalistic phrasing produced fewer ads. Situational requests reached 23.81%, compared with 58.0% for templated lookups.
- Most rendered ads were relevant. Hand review reduced the apparent mismatch rate from 14.3% to roughly 1–1.5%.
- Ad crowding was rare. Two or more ads appeared in 0–5% of responses across most verticals.
- Defensive bidding was almost absent. Only five of 714 rendered ads—0.7%—came from an advertiser also named in the organic recommendation.
Finding 1: Query type was the strongest predictor we found
We first tested four ways of making the same local-business request:
| Phrasing | Ad presence |
|---|---|
| “Best plumbers in Chicago” | 54.3% |
| “Who should I call?” | 58.0% |
| “What is a plumber?” | 58.3% |
| “Plumbers near me” | 62.3% |
All four remained business-recommendation requests. Their confidence intervals overlapped heavily, and the observed rates stayed within a narrow 54–62% band.
We then changed the type of request instead of merely changing the wording. Eight informational legal questions with no city or business-lookup framing were compared with commercial-referral lawyer prompts from the main study.
| Query type | n | Ad presence | 95% Wilson CI |
|---|---|---|---|
| Informational, no lookup framing | 120 | 20.0% | 13.82–28.04% |
| Commercial referral—“find me a lawyer” | 144 | 61.81% | 53.71–69.32% |
The confidence intervals did not overlap.
An initial 40-prompt informational sample produced no attorney-referral ads and looked like a category exclusion. Tripling that sample changed the interpretation. Attorney ads did appear on informational queries, but roughly half as often as self-service legal products such as PublicRecords.us and LegalZoom. The attorneys that appeared were usually matched to the specific issue, such as divorce or custody.
The supported conclusion is not that attorneys are excluded from informational queries. It is that recommendation requests draw substantially more ads, and the advertiser mix changes with the query type.
Finding 2: Local-intent queries produced more ads than the control
The companion citation study had found ads on 59.6% of local-business queries. SE Ranking reported roughly 26% from a single-day snapshot on July 23, 2026. Evertune reported roughly 13% from data collected May 21–June 8, 2026. Both were earlier than our window, and neither methodology disclosed the account tier or login state used for collection.
We therefore built a contemporaneous control using the same collection method.
| Group | n | Ad presence | 95% Wilson CI |
|---|---|---|---|
| Local intent: 10 verticals × 6 markets × 4 templates | 1,200 | 58.25% | 55.44–61.01% |
| General-commercial control: 4 categories, no location | 200 | 40.50% | 33.94–47.42% |
The ratio was 1.44×, and the confidence intervals did not overlap.
The first control sample contained only 100 runs. It produced a 56.00% ad rate with a wide 46.23–65.33% confidence interval, leading us to write the local-versus-general difference up as disconfirmed. After doubling the control to 200 runs, its estimate moved to 40.50%—a 15.5-point change that reversed the conclusion.
Local-intent prompts had a higher observed rate in this window, but that difference did not explain the entire gap from older published baselines. Ad density also appears to have increased across ChatGPT between spring and late summer 2026. The study did not isolate the size of that time effect.
The 1.44× local-intent difference is also smaller than the roughly threefold difference between informational and commercial-referral requests in Finding 1.
Finding 3: Category mattered, but not in the expected order
| Vertical | Ad presence | 95% Wilson CI |
|---|---|---|
| Dentists | 75.83% | 67.45–82.61% |
| Hotels | 70.83% | 62.16–78.22% |
| Lawyers | 68.33% | 59.55–75.98% |
| Restaurants | 68.33% | 59.55–75.98% |
| Plumbers | 67.50% | 58.69–75.22% |
| Florists | 63.33% | 54.42–71.42% |
| Museums | 53.33% | 44.44–62.02% |
| Bookstores | 52.50% | 43.63–61.22% |
| Furniture stores | 44.17% | 35.60–53.10% |
| Auto repair | 18.33% | 12.43–26.20% |
Dentists, lawyers, and plumbers fit a conventional paid-search explanation. Hotels and restaurants were just as high, and florists landed in the middle. Category clearly mattered, but it did not divide neatly into high- and low-commercial-intent groups.
Auto repair was the exception. Its 18.33% rate was less than half the next-lowest category. A separate companion study also placed auto repair last among 12 verticals, at 30.42%—ten points below the next-lowest category.
The measured auto-repair difference was in slots served, not fill rate. Its fill rate—rendered ads divided by served ad slots—was 55.8%, which was not unusually low. But only 0.36 slots were served per run, about one-third of the next-lowest category and one-quarter of the highest categories.
Ad slots served per response were: lawyers 1.23, dentists 1.16, restaurants 1.13, plumbers 1.03, museums 0.96, bookstores 0.89, hotels 0.83, florists 0.80, furniture stores 0.68, and auto repair 0.36.
Some auto-repair ads were loose matches—an appliance-repair company, a car-sharing marketplace, and a solar retailer. That is consistent with limited advertiser coverage. It may represent an opportunity, but this study did not test advertiser demand directly.
Finding 4: Genuinely small towns produced fewer ads
The original market tiers were essentially flat: 56.5–60.75% across dense urban, mid-size, and so-called small metros, with heavily overlapping confidence intervals. Boise and Fargo, however, are regional metros with populations well above 100,000.
We added four genuinely small towns: Hood River, Oregon, about 7,800 people; Sheridan, Wyoming, about 18,000; Marquette, Michigan, about 19,000; and Dodge City, Kansas, about 27,000. Famous resort towns were excluded.
| Market group | n | Ad presence | 95% Wilson CI |
|---|---|---|---|
| Normal markets, same template | 240 | 54.33% | 48.68–59.88% |
| Small towns | 240 | 30.83% | 25.33–36.94% |
An initial 120-run small-town sample produced a 36.67% rate. Doubling the sample moved it to 30.83%. Its confidence interval did not overlap the normal-market interval in either pass.
The difference again appeared in ad supply: 0.49 slots served per small-town run versus 0.80 in normal markets. Fill rates were not worse.
National brands such as Expedia, Hertz, Airbnb, and Wayfair dominated the small-town results. Individually named local businesses appeared less often than they did in normal markets.
The individual town estimates remain imprecise at 60 runs each. Hood River was highest at 38.3%, but its confidence interval overlapped those of the other towns. Only the combined small-town comparison is dependable here.
Finding 5: Naturalistic phrasing produced fewer ads
We wrote one situational prompt for each vertical and compared it with the closest templated prompt from the main study.
| Vertical | Naturalistic | Templated | Gap |
|---|---|---|---|
| Lawyers | 0.0% (n=10) | 76.7% (n=30) | 76.7 points |
| Dentists | 0.0% (n=10) | 70.0% (n=30) | 70.0 points |
| Furniture stores | 0.0% (n=11) | 40.0% (n=30) | 40.0 points |
| Plumbers | 36.4% (n=11) | 73.3% (n=30) | 37.0 points |
| Florists | 30.0% (n=10) | 63.3% (n=30) | 33.3 points |
| Hotels | 40.0% (n=10) | 70.0% (n=30) | 30.0 points |
| Bookstores | 18.2% (n=11) | 43.3% (n=30) | 25.2 points |
| Restaurants | 58.3% (n=12) | 73.3% (n=30) | 15.0 points |
| Auto repair | 10.0% (n=10) | 23.3% (n=30) | 13.3 points |
| Museums | 40.0% (n=10) | 46.7% (n=30) | 6.7 points |
Pooled ad presence was 23.81% for naturalistic prompts and 58.0% for templated prompts, roughly a 2.4× difference.
The individual naturalistic samples were small and their confidence intervals were wide. The direction was nevertheless the same in all ten verticals. Under a one-sided sign-test framing, the probability of all ten moving in the same direction by chance is under 0.1%.
Part of the measured difference was in slots served. Naturalistic requests were served 0.35 ad slots per response, compared with 0.90 for templated requests. Fill rates were similar: 67.6% and 66.4%.
The raw responses showed a second, stronger association with response format. A plain list of business names and addresses appeared in 52% of naturalistic responses. Every one of those responses contained zero ads. Among the other 48%, ad presence was 50%.
The same format appeared in 45% of small-town responses, again with zero ads. It never appeared in the main study’s 1,520 templated, general-commercial, or informational-query runs.
This is the smallest and most exploratory comparison in the article: ten handwritten scenarios with ten or more repetitions each. It suggests that realistic situational phrasing is associated with both fewer ads and a different response format. It does not establish how often real users phrase requests this way.
Who bought the placements
The leading advertisers in the local-intent group were Expedia (69), Thumbtack (53), UrbanStems (51), Priceline (32), GoTickets (24), and DoorDash (22), followed by named dental and legal practices, Wayfair, trivago, and Grubhub.
The general-commercial group was led by BestMoney, Insurify, Rocket Mortgage, Fetch Pet Insurance, SoFi, LegalZoom, and Liberty Home Guard. No individual local businesses appeared in that control group.
Local-intent results included some individually named local businesses, but travel and delivery brands outweighed Thumbtack and the named local businesses combined. Local intent changed the advertiser mix; it did not make that mix predominantly local.
Ad quality: usually relevant, rarely crowded, almost never defensive
Relevance
A keyword heuristic initially flagged 14.3% of ads as mismatched. Hand review showed that most flags came from unrecognized brand names—such as Marriott, Pottery Barn, Philips Sonicare, and named law firms—or reasonable cross-selling, such as Expedia and AllTrails on museum queries.
After review, an estimated 7–10 of 714 rendered ads were genuinely mismatched: roughly 1–1.5%. Examples included a pet clinic on dentist queries and an online K–12 school on a museum query.
Crowding
Two or more ads appeared in 0–5% of responses across most verticals. Lawyers were highest at 5.0%.
Defensive bidding
Only five of 714 rendered ads—0.7%—came from an advertiser also named in the response’s organic recommendation. All five involved large national brands such as UrbanStems and Expedia. No small local business in the sample appeared to be defending its own organic mention.
Predictions, scored
| Prediction | Prior confidence | Result |
|---|---|---|
| Ad presence correlates with commercial intent | ~65% | Not supported by small wording changes; supported by the informational-versus-referral comparison, which moved 42 points |
| Local-intent ad rate is at least 1.5× the general control | ~80% | Close but below the threshold: 1.44× after expanding the control; the initial estimate was 1.04× |
| Paid-search-like categories score higher | ~70% | Partially supported; the gradient did not split cleanly, and auto repair was a replicated low outlier |
| Local queries attract more local lead-generation advertisers | ~60% | Partially supported; local businesses appeared, but national travel and delivery brands outweighed them |
| The collection reflects Free-tier-equivalent exposure | ~50% | Supported by a logged-in Free-tier browser check and written confirmation from DataForSEO |
The lesson from the control group
Two headlin
See the stats/study in context: Steady Demand Research Index
