Three things happened in sequence, and I've written about each of them.
First, the click died. Roughly 68% of Google searches now end without a click to any website, and when an AI Overview appears above the results, click-through to the pages below drops by more than a third. The search result page stopped being a doorway and became a destination.
Second, the question moved. Pet owners started asking assistants directly — not typing "emergency vet open now" into a search box, but asking ChatGPT whether the symptoms they're describing sound urgent, and where to go. OpenAI noticed, and started testing ads inside those conversations in February.
The third step hasn't fully happened yet, and it's the subject of this post: the assistant stops answering and starts acting. Not "here are three cat clinics" but "I found a feline-only practice with an opening Thursday at 2:40 — want me to book it?"
When that happens, the referral — the moment a pet owner gets pointed at your practice — is made by a model. The referring vet is an AI.
I have opinions about what that means for an independent practice, and opinions are cheap. So here are four predictions with dates attached. Next August I'll publish a scoring post and grade each one — right, wrong, or partial — with evidence. If I'm wrong, you'll be able to quote me on it.
Prediction one: by August 2027, at least one major assistant answers "find me a vet" with a shortlist and a booking action.
Not a list of links. Not a map handoff. A recommendation you can act on inside the conversation — in production, in ordinary US metros, for ordinary local queries.
The mechanism is already visible. The assistant platforms are all building agent capabilities that fill forms and complete transactions, and local services are the obvious proving ground: high intent, clear action, measurable outcome. Veterinary care sits in the sweet spot — urgent enough that people want the friction gone, routine enough that a booking is a booking.
Confidence: high. What would prove me wrong: a year from now, every major assistant still ends the local-vet conversation by handing you a list of links or bouncing you to a maps product, with no action attached.
Prediction two: by August 2027, review text will matter more than review stars in assistant recommendations.
A language model doesn't read your Google profile the way the old ranking systems did. It reads your reviews as text. A practice at 4.6 with four hundred reviews that specifically mention a blocked cat handled at 2 a.m., a fearful dog treated patiently, a dental estimate that matched the invoice — that practice gives a model something to reason with. A 4.9 built on forty reviews that say "great vet, highly recommend" gives it almost nothing.
This inverts a decade of review strategy. The advice was: maximize the average, maximize the count. The new version is: the content of the review corpus is the asset. Specific, situational, condition-naming reviews are what an assistant can match against a pet owner's described problem.
Confidence: moderate. This one is hard to measure from the outside, which is part of why I'm committing to it now — I'll have to get creative about scoring it honestly. What would prove me wrong: systematic testing next summer shows assistant recommendations tracking star averages and map rank, indifferent to what the reviews actually say.
Prediction three: my "not yet" on ChatGPT ads holds through at least mid-2027.
I published the verdict last week: conversational ads aren't ready for a single-practice budget, because you can't verify the local targeting and you can't measure conversions against your own site. I'm now betting that stays true for another year — the ad product will get more polished, possibly add cleaner geo tiers, and still lack the thing that matters: evidence, in your own analytics, that a dollar in produced a booked patient out.
The mechanism behind this prediction is structural, not technical. There's no conversation-level reporting because the conversations are the product's privacy promise. Building advertiser-grade measurement means deciding how much of that promise to spend, and that's a slow, reluctant decision — not a feature sprint.
Confidence: moderate-high. What would prove me wrong — and I'd be glad to be wrong — is verifiable metro-level targeting plus own-site conversion import shipping before next summer. The day that happens, my watch-list post becomes a how-to post.
Prediction four: within two years, "AI visibility" becomes auditable the way page speed is — and most practices fail it.
Today I can run any practice's site through PageSpeed Insights and show the owner a number. In two years, I expect the equivalent exists for assistant-readiness, whether as a formal tool or as a repeatable checklist: Is your practice's structured data complete and correct — services, species, hours, emergency status — in a form a model can consume? Can an agent actually book with you, or does your "online scheduling" end in a callback form? Does your review corpus read like evidence or like filler? When someone asks an assistant about your niche in your city, do you appear at all?
Most independent practice sites will fail that audit the way they fail Core Web Vitals today: not because anyone decided to fail, but because nobody was measuring. The practices that pass will be disproportionately the ones that own their site and their data outright — which is not a neutral observation coming from me, since ownership is the way I build. But the logic stands on its own: you can't make your data machine-readable if a vendor holds it hostage.
Confidence: high on direction, honest uncertainty on the timeline. What would prove me wrong: assistant recommendations turn out to be a thin wrapper on existing map rankings, structured data goes unread, and the whole audit reduces to "do your old SEO."
What survives
Strip the four predictions down and they share one skeleton. When the referring vet is an AI, three things hold their value:
Verifiable data. Structured, accurate, machine-readable facts about what you treat, when you're open, and what you charge. Models recommend what they can verify.
Bookable inventory. An appointment an agent can actually take. If acting on your practice requires a phone call and a hold queue, the agent recommends the practice next door.
Paid presence in the answer. When the ad products mature past my prediction three, the auction moves inside the conversation. The practices that already measure — which dollar produced which patient — will be the ones that can afford to bid.
Notice what's not on the list: your homepage hero image, your position on a results page nobody scrolls, your star average to one decimal place. The click already died. The question already moved. The predictions above are just the bill arriving.
I'll see you back here in August 2027 with the scorecard.
These are predictions, not certainties — each one states its confidence and what would falsify it, and the scoring post will grade them against evidence, not vibes.