When a machine recommends your practice, it is reading your reviews

BrightLocal published its 2026 Local Consumer Review Survey on 11 February, from a representative panel of 1,002 US adults run through SurveyMonkey. Two of its findings belong together, and they are usually quoted apart.

The first is that Google slipped. BrightLocal has asked the where-do-you-read-reviews question for years and Google has always been the standout answer. This year its share dipped from 83% in 2025 to 71%. BrightLocal is clear that this is not a migration to AI — Facebook, Tripadvisor, Apple Maps and Healthgrades all grew too, and it notes that only 35% of small businesses even have a Google profile. So take the second finding without a causal arrow between them. BrightLocal’s 2026 survey puts use of ChatGPT and other generative AI tools for local recommendations at 45%, up from 6% the year before, which in their data makes it the third most common recommendation source behind Google and Facebook. Alongside that, 40% of respondents said they trust AI platforms to provide business recommendations, and 42% said they trust AI platforms as much as they trust traditional reviews.

I want to be careful with that 6%-to-45% figure. It is a very large single-year move, it is self-reported, and it comes from one survey of a thousand people. People are unreliable witnesses to their own habits, and “have you used AI to find a local business?” is exactly the sort of question that picks up a year of headlines as readily as a year of behaviour. Read it as a direction rather than a measurement. The same caveat applies to everything else in the survey, including the two numbers I am about to use: 82% of consumers say they read the AI summaries that sit above the reviews on a listing, and 23% say they would decide on that summary alone. That is not an assistant recommending you. It is the same artefact though — a machine writing a paragraph about you, out of your reviews.

That last number is the one I would sit with. A summary is not a list. A list shows a patient your stars and your count and lets them judge. A summary shows them a sentence that somebody else wrote about you.

The assistant cannot see your practice

When someone asks an assistant for the best pediatric dentist near them, the assistant has no view of your operatory, your staff, your chairside manner or your sterilisation logs. It has text about you. For a hospital system that text includes news coverage, research output and a well-maintained site. For an ordinary local practice there is very little else in writing. No news coverage, no research output, usually a thin site. Whatever else an assistant reads about you, your reviews are the bulk of what exists to read — which is a claim about supply, not about how any model weights it.

Three failure modes follow, and I would guess most practices reading this are in one of them.

Too little text. A practice with eleven reviews gives a summariser almost nothing to work with. Anything it says about you is drawn from a handful of sentences, so whatever it says about you is drawn from a handful of sentences. I have watched it hedge, and I have watched it name a practice down the road that it could actually describe. I cannot tell you which it will do on any given run. This is not a rare corner case. In our own 2026 dataset of 15,061,473 Google reviews across 142,417 US medical practices, the median practice holds 19 reviews and 38.5% hold fewer than ten.

Stale text.Recency was already load-bearing for human readers — BrightLocal puts the share of consumers seeking reviews written in the last three months at 74%. A model reading a corpus whose newest entry is from two years ago is describing a practice as it was, if it describes it at all.

Empty text.A practice with 300 reviews that all say “great experience” and “highly recommend” is barely better positioned than the practice with eleven. There is no distinguishing language in it. Ask a summariser to characterise 300 interchangeable compliments and it will give you back one interchangeable compliment. My working assumption — and it is an assumption — is that the practices described well are the ones whose reviews contain specifics: the procedure, the anxiety that got handled, how long the wait actually was, whether the billing was explained before or after, the name of the hygienist.

Google has documented its side of this, and it has not moved

Google’s own help page: “Local results are mainly based on relevance, distance, and popularity.” Its section on that third one — headed Prominence — says on reviews, plainly, “More reviews and positive ratings can help your business’s local ranking.” And on the shortcuts: “There’s no way to request or pay for a better local ranking on Google.”

That has been true for years and it stays true. The levers have not changed. What has changed is the audience for them. Google names two things about reviews: how many you have, and whether the ratings are positive. Recency and wording are not in that sentence. They are, however, most of what a summariser has to work with. Count and rating are enough to place you in a list. Neither can be turned into a paragraph of prose about you. If you want the mechanics of the ranking side, we broke those down separately in the Local Pack factor breakdown; this post is about the other consumer of the same data.

Run the query on yourself, this week

Here is the one thing I would actually do, and it takes about twenty minutes.

Open a generative assistant logged out, or in a temporary chat with memory off — your own account has spent months talking to you about your practice and will happily tell you what you want to hear — and ask it the three or four phrasings a real patient would use for your specialty and your area. Not your marketing phrasing. The way a worried parent types at nine at night: “best pediatric dentist near me”, “dentist in Austin good with anxious kids”, “who should I take my 4 year old to for a filling in Austin”. Write down three things.

1. Whether you are named at all. Binary. No interpretation required.
2. Which competitors are named. Note the ones you did not expect.
3. The exact adjectives it uses about each of you. Copy them out word for word. This is the finding.
4. Whether the answer is even stable. Run each phrasing twice, in two separate chats. If the two runs name different practices, that is a finding too: nothing about you is settled enough in the text for the answer to converge.

The third line is where the diagnosis lives. If the assistant describes a competitor as “known for short wait times and clear pricing” and describes you as “a dental practice in Austin”, you have not found a ranking problem. You have found a text problem. Their reviews contain sentences and yours contain ratings. No amount of profile optimisation closes that gap, because the gap is in what patients wrote.

Which turns the fix into a question most practices have never treated as one: what exactly are you asking patients, and does the ask invite a sentence or a star? “Please leave us a 5-star review” is an instruction to produce a rating, and patients comply with it exactly. “How did the appointment go?” is an opening. This is survey design, not marketing. The wording of the question determines the shape of the answer, and the shape of the answer is the asset.

The ask decides the answer

This is the thing applaud was built to do, so I will state it once and plainly. Robin, our agent, has a conversation with the patient after a visit instead of sending a form — it texts and calls in the practice’s own name and number, and patients reply in their own words. That is what produces sentences rather than scores. Every patient who visited is enrolled, not a hand-picked list, so the asks go out at the rate patients actually walk through the door rather than in campaign bursts. And the review link opens directly in Google’s composer, the box with the stars, ready to type — because the friction between “yes I would” and a posted review is where a review programme loses volume, and it loses it silently.

Every patient is invited to review; what a patient says never removes that invitation. What it changes is sequencing. A frustrated patient goes to a human first — the complaint lands in the practice’s queue with the team notified, and the queued automated ask is held until it is dealt with, not cancelled. The survey half is free, and you pay when a review posts.

The part I cannot tell you

Nobody outside the labs building these models knows how the assistants weight what they read, and that includes us. I can tell you what is in the corpus. I cannot tell you the function applied to it, whether it is stable between model versions, or how much of the answer is coming from reviews versus a directory listing versus something Google fed the model directly. Anyone who gives you a percentage breakdown of that is guessing.

The survey is self-reported intent, not observed behaviour. And a practice that reorganises its entire marketing around an AI answer box in 2026 is making a bet on a distribution channel that did not exist three years ago and may look different in two.

The reason to act anyway is that the work is the same work. More reviews, arriving more recently, containing actual sentences about what happened. A patient scanning a list wants that. Google says it helps. And it is the only thing a summariser can build a sentence out of. You do not have to believe the 45% to do it. If the bet is wrong you have still built the thing you should have built for the ordinary case, which is the only version of this argument I am willing to make.

Sources

1. BrightLocal. “Local Consumer Review Survey 2026.” Published 11 February 2026; representative panel of 1,002 US adults via SurveyMonkey. brightlocal.com/research/local-consumer-review-survey. “Use of ChatGPT and other generative AI tools for local recommendations has grown rapidly, rising from 6% last year to 45% and becoming the third most popular source of business recommendations.” Google’s share as a review source “dipped from 83% in 2025 to 71%”. 40% trust AI platforms to provide business recommendations; 42% trust AI platforms as much as traditional reviews; 82% read AI-generated review summaries, 23% willing to rely solely on them; 74% seek reviews written in the last three months.
2. Google. “How local results rank on Google Search and Maps.” Google Business Profile Help. support.google.com/business/answer/7091. Relevance, distance and prominence. “More reviews and positive ratings can help your business’s local ranking.” “There’s no way to request or pay for a better local ranking on Google.”
3. applaud. “The State of Patient Reviews in American Medicine.” 2026. applaud.you/newsroom/state-of-patient-reviews-in-american-medicine. 15,061,473 Google reviews across 142,417 US medical practices; median practice holds 19 reviews; 38.5% hold fewer than ten.