We read 404 unique insurance creatives from 101 advertisers across three public ad libraries. Carriers wrote most of them. Comparison sites and media brands took the brand-name prompts. A quarter of the prompts that triggered an insurance ad never used the word insurance.
You can’t see your competitors’ ChatGPT ads in Ads Manager. You can’t see the prompts either.
So we went where you can. Three public libraries record ChatGPT ads by asking it questions and saving what shows up. We pulled every creative tagged insurance, removed the duplicates, and read them one by one. Here’s what 101 insurance advertisers are doing right now.
Who is advertising?
404 unique creatives from 101 advertisers, classified by hand. The libraries: sprites.ai, chatgptadlibrary.com and gptadlibrary.com.
| Advertiser type | Creatives | What they write |
|---|---|---|
| Carriers | 157 | Benefit lines: “Bundle Policies and Save With GEICO”, “Reimburse Vet Bills”, “Get a quote in a few taps” |
| Comparison sites | 112 | Match lines: “Insurance near you, find local providers ready to help” (EverQuote, shown 482 times) |
| Media brands | 80 | List lines: “Best 5 Value Pet Plans 2026”, “10 Best Auto & Home Insurance 2026” |
| Agencies | 25 | Local and line-specific offers |
| B2B and other | 11 | Certificates of insurance, commercial lines |
Scale is concentrated. On gptadlibrary’s October capture, EverQuote (933 ads shown), GEICO (510) and The Zebra (472) took 31.5% of the 6,075 insurance ads recorded. GEICO alone runs 45 distinct copies. Allianz, New York Life, Forbes, NerdWallet and BestMoney fill out the top ten.
A regional carrier? Those are the cards you are bidding against.
What do the prompts look like?
Nothing like search queries.
gptadlibrary classifies the prompts that triggered insurance ads as “best X for me” 20.2%, comparisons 15.0%, price 15.1%, long personal scenarios 14.9%, “fix or switch” 14.7%, “is it worth it” 10.1%, short searches 10.0%. Across the 1,279 unique prompts in our sample, our own tagging found 267 best-for or cheapest prompts, 233 comparisons, 228 price prompts and 160 long scenarios of 18 words or more.
Study the long ones. “I’m 27, got my license in Ohio last year and had one minor fender bender already.” “Lemonade auto home pet bundle, is it really cheaper or just marketed that way.” “Is Lemonade renters insurance legit, Reddit.”
That last one returned a Lemonade ad: “Best Renters Insurance, starting at $5/mo.” A trust question got a price answer. That is a mismatch you can beat.
Who wins when someone names a brand?
221 of the 1,279 prompts named an insurer. On those, the named brand’s own ad appeared 23 times. A rival carrier appeared 68 times. A comparison site or media brand, 140 times.
GEICO is the main conqueror: 12 of 28 State Farm prompts, 8 of 35 Progressive prompts, 5 of 10 USAA prompts, 4 of 13 Allstate prompts. BestMoney is the most frequent third party on every large brand.
Two lessons. Big brand? Your name is a prompt your rivals describe in their hints. Describe it in yours. Small brand? The comparison prompts (“state farm vs allstate auto”) are open to anyone who writes a better card than a listicle.
The prompts with no insurance word
312 of the 1,279 prompts contained no insurance word at all.
Liberty Mutual’s “Switch and Save” appeared on “best phone deals 2026 with trade in.” GEICO appeared on “is Valvoline worth it for someone who drives 15,000 miles a year.” Lemonade’s pet ads appeared on dog-vet and cat-feeder questions. ERGO NEXT’s general-liability certificate ad appeared on a question about cannabis-compliant soil testing labs.
Some of those are good matches. A 15,000-mile driver is an auto prospect. Some are noise. Either way they prove the mechanism: the platform reads the situation, not the word. Write hints for the situation. Chapter 4 shows how.
How do they write, and what is on the card?
Median headline: 25 characters. Median description: 54. 34% of creatives use a comparison word (compare, vs, best, top). 11.4% mention a price. 8.4% ask a question. Media brands put the year in the headline; carriers do not. Nearly every card carries an image: 95% on gptadlibrary, 99% on chatgptadlibrary.
Now set those medians beside the cut a phone makes at about 24 and 48 characters. Most of the industry is losing the last word of its headline and the last clause of its description. That is the cheapest fix in this guide.
What we saw live, and how NEXT runs it
Libraries are secondhand. So on October 27 we opened ChatGPT signed out, on a phone-sized screen, and asked three questions a small-business owner would ask.
- “What does cyber insurance actually cover for a small business?” Ad: Hiscox, “Need Business Insurance? Get coverage built for your business needs.”
- “Is it required in North Carolina for an LLC? We’re a 12 person marketing agency in Wilmington.” Ad: ERGO NEXT Insurance, “Small Business Insurance. Instant quotes from top local brokers online,” with a “starting from $19/month” panel.
- “Who offers the best rates for a company our size, and should I go with a national carrier or a local agent?” Ad: the same ERGO NEXT card.
Three questions, three ads, on the first message each time. Meanwhile the answer itself named Hiscox, The Hartford, Chubb and Coalition, quoted Insureon’s premium data, and listed two Wilmington brokers with street addresses. Being named in the answer and buying the card are two different games, and NEXT chose the card.
The click told us more than the card. NEXT’s landing URL carries its campaign structure in plain text:
| In the URL | What it tells you |
|---|---|
/lps/business-insurance-oa/ | A dedicated landing page, not the homepage. We read “oa” as OpenAI; NEXT has not said. |
utm_source=chatgpt, channel=sem | ChatGPT tracked as its own source, bucketed with search in their reporting. |
utm_campaign=Generic-Business-Insurance-CPA-CPC_billing_Lead | A generic business-insurance campaign, conversions objective (CPA), billed on clicks, optimising to a lead event. Exactly the setup in Chapter 5. |
adgroup=Cost-Pricing-Business-CPA-CPC_billing_Lead | The ad group is an intent theme: cost and pricing questions. Our “who offers the best rates” prompt landed in it. |
utm_content=20 | Ad variant 20. They are testing a lot of cards. |
oppref=, olref= | OpenAI’s click references, the macros from the developer docs, so the server event can be matched back to the click. |
tracking_id= | NEXT’s own ID on top, for its CRM. |
The page the click lands on is a business-insurance page with an industry picker (construction, food and beverage, retail, cleaning), a 4.7 out of 5 rating from 5,056 reviews, “750,000+” customers, partner logos, four awards, and the same “Prices start at just $19/month for general liability” line the ad promised, with a footnote that says starting prices are existing premiums as of April 1, 2026 and not guaranteed. Every button goes to the quote flow. The page never mentions ChatGPT. That is a landing page that keeps the ad’s promise, and it is the one I would copy the shape of.
Method, and what this data cannot tell you
These are probe libraries. Researchers write the prompts. On gptadlibrary, each ad appeared after a follow-up message asking for a sponsored ad, and the site itself says its counts show who advertises, not how often ChatGPT shows ads.
The libraries also gate part of their data. chatgptadlibrary shows 24 of its 50 insurance ads without a sign-in. sprites shows 12 per advertiser. The dates differ: sprites June to August, chatgptadlibrary June 19 to August 3, gptadlibrary October 3 to 4. So this is a large sample of who is running what. Not a census. Not a measure of frequency or performance.
One more honest line: none of this tells you which ads worked. The libraries record appearances, not clicks. The performance evidence, such as it is, lives in Chapter 7.
FAQs
Can I republish the ads from these libraries?
Don’t. The creatives belong to the advertisers and the libraries have their own terms. Read them, count them, learn from them, write your own.
Is there an official ad library from OpenAI?
Not as of today. These three independent libraries and the Brookings dataset are what exists.
How often do insurance prompts show an ad?
Nobody has a clean number. Studies range from 4.5% of prompts to 49% in the US, and one Australian vendor reported 71% on insurance queries over two days. Insurance seems to draw ads more than average. That is as far as the evidence goes.
Sources, checked October 26, 2026:
- sprites.ai, insurance industry page
- chatgptadlibrary.com (Finance > Insurance and the sub-niches)
- gptadlibrary.com (insurance industry section, October 3 to 4, 2026 capture)
- Vital Design, two creative tactics (the 24 and 48 cut)
