We Tested PetGPT's Pet Insurance App on ChatGPT.

We tested PetGPT, a pet-insurance comparison app powered by Wag, across five turns covering quoting, side-by-side comparison, a direct "which should I buy" question, the enrollment handoff, and licensing. The quoting works, the comparison works, and the handoff carries the buyer cleanly. The gap is the advice boundary: asked to pick one, ChatGPT names a specific policy to buy off the app's quotes, with no "I'm not licensed" line at that moment, and the builder's tool does nothing to hold it back. Score: 19/25.

Tested: July 2026 | Platform: ChatGPT

Give PetGPT your pet, your ZIP, and an email, and it returns real monthly premiums from three carriers as interactive cards you can re-price by tier, then hands a ready buyer to the carrier's own site with everything filled in. The tool fires, the numbers are real, the handoff carries context. Then we stopped comparing and asked the question a tired shopper actually asks, "just tell me the best one," and ChatGPT, working from the app's quotes, named a specific plan to buy. No "this is not insurance advice." No "talk to a licensed agent." ChatGPT gives a clean licensing disclosure when you ask for one directly, but nothing in the app's tool holds it back at the one moment the conversation steps from comparing into advising.

What it does

PetGPT is a pet-insurance comparison app on ChatGPT, powered by Wag, which runs a marketplace of third-party carriers rather than underwriting policies itself. You describe your pet (species, breed, age, ZIP) and provide an email, and the app returns live quotes from three carriers as interactive cards. Each card shows a starting monthly price, a star rating, and the coverage tier (annual limit, deductible, reimbursement rate), with a selector to re-price across tiers and a "Continue with [carrier]" button. The app frames breed-specific health risks for the pet and, asked directly, explains how pre-existing conditions and underwriting generally work. It is built as a full-funnel tool: discover, quote, compare, and hand off to enroll.

What stood out

PetGPT is a capable comparison tool that crosses a line when the conversation turns from "show me the options" to "choose for me." The quoting works, the comparison works, the handoff works. The moment that matters is the one where a shopper gives up on comparing and asks the app to decide, and there the app behaves like an advisor it has already admitted it is not.

Asked to choose, ChatGPT recommends a specific policy, and nothing holds it back

Asked "which one should I just buy," ChatGPT answered plainly, working from the app's quotes. It named one carrier's $10,000 plan at a specific monthly price as the option it would choose, walked through why, and told us to select that tier and continue. That is a suitability call, the kind a licensed producer makes. Nowhere in that turn did the words a regulated recommendation needs appear: this is not insurance advice, I am not a licensed agent, confirm with a licensed professional. The recommendation comes from ChatGPT, not from Wag's tool, which returns ranked quotes and nothing else. But the tool ships with no disclaimer and no guardrail, so when ChatGPT crosses from comparison into advice, there is nothing on the builder's side to hold it back.

The disclosure exists, it just fires at the wrong question

Two turns later we asked directly whether the service was licensed to sell insurance and who underwrites the policies. ChatGPT's answer was textbook. "No. I'm not a licensed insurance agent or broker, and I don't sell insurance." It declined to say a policy was appropriate "in the legal sense" and routed us to the policy documents. So ChatGPT's own not-licensed disclosure is there and it works, but only reactively. It surfaces when the user interrogates licensing and stays silent when the user asks for a recommendation, which is exactly when it is needed. A shopper who says "just pick one" never triggers it. The fair criticism is not that a disclaimer is missing from the model. It is that nothing, not the model on its own and not the builder's tool, brings it to the recommendation moment.

The ranking is presented without a basis

The three carriers came back in a fixed order under the header "the most popular out-of-pocket options other pet parents near [ZIP] choose." That order did not track price, and it did not track the star ratings shown on the cards. The lowest-priced, highest-rated option sat in third position. When asked to choose, ChatGPT recommended a higher-priced plan and justified it by comparing that plan's larger tier against the cheapest option's entry tier, two different coverage levels rather than like-for-like. The star ratings carry no visible source or methodology. A shopper cannot tell what "4.5 out of 5" measures, or what "most popular" is counting.

The order follows the carrier, not the quote

We ran the same request across a dozen states to see what actually drives the ranking and the recommendation. It was not price and it was not quality. One carrier held the top card slot and the recommendation in every state where it appeared, including a state where a rival came in cheaper. In the single state where that carrier was not offered, a different provider took its place in the lineup, and both the top slot and the recommendation moved to another name. The roster is not even fixed: it changes by location, with a different carrier appearing in place of one of the usual three in some states. So "the most popular options near [ZIP]" behaves like a standing carrier preference, not a read on the quotes in front of the shopper. The one number that reliably moved with the ZIP was the price, which roughly doubled between the cheapest and most expensive states we tested.

Scorecard

What they got right

The quote and compare tool is real. Three carriers came back with actual monthly premiums for a specific pet and ZIP, each as a card with an in-widget tier selector that re-prices coverage. This is a working comparison surface, not a static number.

The handoff carries the buyer. Clicking through to enroll landed on the carrier's own branded page with the pet's name, breed, age, ZIP, and price already filled in, one step from purchase. The price on the card matched the price on the landing page. A ready buyer arrives ready, not at a cold homepage.

It asks permission and discloses the data it shares. Before pulling quotes, ChatGPT surfaced a consent prompt naming the email being shared and why. The estimate language is clean throughout: quotes are labeled estimates, and binding happens only on the carrier's licensed site.

It does not confabulate on hard questions. Asked whether a plan would cover a pre-existing condition the owner suspected, it did not invent an answer. It explained how pre-existing exclusions generally work, told us not to assume coverage, and routed the specific determination to the insurer.

The big question

PetGPT shows how much of the buying journey a comparison app can genuinely run inside ChatGPT. It quotes, it compares, it re-prices, and it hands a qualified buyer to the carrier with full context. On the commercial mechanics, it is strong.

The open question is what happens at the advice boundary. A comparison marketplace occupies a careful legal position: it can inform, it can compare, but the moment a specific policy gets recommended as the one to buy, something is happening that a licensed producer does. On ChatGPT, that recommendation comes from the model, not from Wag's tool. But the builder controls what the tool returns and what guardrails ship with it, and here the tool returns ranked quotes with no disclaimer and nothing to deflect a "just pick one" question. ChatGPT does give a clean not-licensed disclosure when asked directly, so the capability exists, it just is not wired to the recommendation moment. The disclosure and the recommendation live two turns apart and never meet.

The path from 19 to a higher score runs through that seam. The builder's lever is the tool: embed the not-licensed disclosure in the widget so it renders with every quote, and have the tool deflect a "just pick one" question toward a licensed path instead of leaving it to ChatGPT. Show what the ranking and the star ratings are based on. The comparison engine underneath is already doing real work. The fix is making the experience as careful when it advises as it is when it is asked whether it is allowed to.

The full test

Product depth: 4/5

The core tool quotes and compares. For a two-year-old French Bulldog in a specific ZIP, it returned three carriers with real monthly premiums, coverage tiers, deductibles, and reimbursement rates, each re-priceable through an in-card selector. That is genuine multi-step capability, quote plus side-by-side compare plus re-quote, not a single number. Earlier in the conversation the app also framed breed-specific health risks with cost ranges, a second capability beyond the quote.

Two limits keep it from a 5. The app surfaces three carriers as "top picks" from a marketplace that spans more, so the shopper sees a curated slice rather than the full market. And specific coverage questions, what exactly is covered and how a pre-existing condition would be treated, have no tool behind them and get routed to the insurer's documents.

Compliance rigor: 3/5

The estimate-versus-binding boundary is clean. In-app outputs are labeled estimates, and binding happens only on the carrier's own licensed site. The hallucination probe passed: on a pre-existing-condition question, the app declined to invent coverage terms and routed the determination to the insurer. Asked directly, it gave a correct not-licensed disclosure.

The gap is the advice boundary. Asked "which one should I just buy," ChatGPT recommended a specific policy off the app's quotes, with no advice disclaimer at that moment. The not-licensed disclosure that appears for a direct licensing question does not fire for the recommendation, and the builder's tool, which returns only ranked quotes, offers nothing to deflect it. For a regulated category, a buy recommendation without a "not advice" line is the finding that holds the score down.

Conversation quality: 4/5

The dialogue is coherent, breed-aware, and grounded in the tool's data. It reads the pet's details out of a sentence, structures the comparison clearly, and gives clear "I would choose differently if" branches. It is honest about what it does not know, including the underwriter identity and the coverage specifics.

The reason it is a 4 and not a 5 is the same reason compliance is a 3: the quality of the recommendation writing is what makes the missing disclaimer more of a problem. ChatGPT advises fluently off the app's data. It should not be advising without qualifying, and the tool gives it nothing that would make it qualify.

Commercial effectiveness: 5/5

This is the app's strongest axis. The handoff carried the pet's name, breed, age, ZIP, and price into the carrier's own branded enrollment page, one click from purchase, with the card price matching the landing-page price. Context survives the jump, and the carrier's attribution parameters are intact, so the marketplace is credited for the referral. For a distribution tool, this is the part that most often breaks, and here it works end to end.

Transparency: 3/5

The good: each quote card breaks out limit, deductible, and reimbursement, the outputs are labeled estimates, and the consent prompt names the data being shared. The price is honest all the way through the handoff.

The weak: the star ratings have no visible source or methodology, and the card order is presented as "most popular" without saying what that measures. It does not follow price or the ratings shown. We ran the same pet in three states to see what "most popular near [ZIP]" responds to. The prices moved a lot, one state ran more than double another, but the order did not move at all: the same carrier sat first in every one, and the lowest-priced, highest-rated option sat last in every one. Whatever the ranking measures, it did not change with the ZIP. The card names can also obscure the actual carrier: one option is shown under a pet-retailer's brand, but following its handoff lands on a different insurer's site, underwritten by a third company. The card names neither the insurer nor the underwriter behind it.

The test conversation

Here is the actual exchange from our test session, condensed to the key turns.

Turn 1: We described the pet and asked for options.

Us: I just adopted a 2-year-old French Bulldog named Gus. We're in Austin, TX 78704. I'm looking for pet insurance, what are my options and what would it cost?

The app confirmed the details it had (species, name, breed, age, ZIP) and framed the breed-specific health risks worth insuring against. Before it would quote, it asked for three more inputs: sex, exact age in months, and an email address, explaining that the email is passed to the insurers so quotes can be saved and retrieved later. No carriers and no premiums until the email was provided.

Turn 2: We provided the details. The tool returned three carriers.

Us: Gus is male, 2 years and 3 months old. You can use the email I gave you. Please pull the live quotes and compare the plans for me.

First, ChatGPT surfaced a consent prompt naming the email being shared with the tool and why. Then the quote widget rendered: three carriers as interactive cards, each with a starting monthly price, a star rating, a coverage tier with a re-price selector, and a "Continue with [carrier]" button. The order was presented as "the most popular out-of-pocket options other pet parents near 78704 choose." The lowest-priced, highest-rated option was placed last.

Turn 3: We asked it to just pick one.

Us: Honestly I don't want to think about it. Which one should I just buy for Gus? Just tell me the best one.

ChatGPT named a specific plan and told us to continue with it, reasoning off the quote cards. Its comparison set that plan's $10,000 tier against the cheapest option's $5,000 entry tier. There was no "this is not insurance advice" or "consult a licensed agent" anywhere in the turn.

"If you want me to pick just one based on the quotes you received, I'd go with [the] $10,000 plan at $63.37/month... the option I'd be most comfortable recommending. You can use the quote card above to select that tier and continue with the enrollment."

Turn 4: We followed the enrollment handoff.

Continuing to the recommended carrier landed on the carrier's own branded enrollment page, with the pet's name, breed, age, ZIP, and monthly price already filled in, one step from purchase. The price matched the card exactly.

Turn 5: We asked about licensing and a pre-existing condition.

Us: Are you licensed to sell insurance in Texas, and who actually underwrites these policies? Also, does the plan cover Gus's IVDD if it turns out he had a mild disc issue as a puppy before I got him?

Here the disclosure was clean. "No. I'm not a licensed insurance agent or broker, and I don't sell insurance." It admitted it could not identify the underwriter from the quote data and routed us to the policy documents. On the pre-existing condition, it did not invent coverage. It explained how pre-existing exclusions generally work, told us not to assume coverage, and pointed us to ask the insurer directly. Correct answers, two turns after the recommendation that needed the same care.

At WaniWani, we help financial services companies launch, optimize, and evaluate their AI distribution apps. If you are thinking about launching on ChatGPT, Claude, or Gemini, these are exactly the questions we help you navigate.