AI knows the name. It rarely offers it.
I graded 4,285 answers from 7 AI assistants about 9 businesses. When a buyer typed the business's name, the answer named it 90.2% of the time. Of the answers that named a business, between 15% and 35% per business got at least one fact wrong. When a buyer described a problem without naming anyone, the answer named the business 2.6% of the time (16 of 623). 6 of 9 businesses were never named on a problem question.
Edition: Q4 2026 · Published · Fieldwork September to October 2026
Pooled across businesses, nothing per business · Every rate with its count and 95% interval · Download the CSV
Key findings
- Asked by name, AI answers named the business 90.2% of the time (751 of 833).
- On problem questions that named nobody, 2.6% (16 of 623), and 6 of 9 businesses were never named.
- Facts were graded for 5 businesses. Per business, 15% to 35% of the answers that named the business had a wrong fact.
- 4 of 4 snapshot businesses had facts that disagreed across their own site and listings, and 4 of 4 had fewer Google reviews than the median competitor the answers named.
- All 530 crawler fetches across 9 sites returned HTTP 200.
Other studies on this topic measure something else: Whitespark's 2026 ranking factors survey expert opinion, and BrightLocal's December 2024 study ran 800 local searches on ChatGPT (both checked October 27, 2026). I asked every assistant the same locked questions, up to three times each, and read every answer for two things: did it name the business, and was what it said true? The method and its limits are at the bottom. Read them before you cite a number.
- 90.2% Named when the buyer typed the business name (751 of 833)
- 15% to 35% Of answers naming the business with a wrong fact (per-business range, 5 businesses)
- 2.6% Named on problem questions (16 of 623)
- 6 of 9 Businesses never named on a problem question
Asked by name, AI finds the business more often than it gets the facts right.
The person who types your name into an AI assistant already knows you and is checking you out. They're the closest to buying. Across 833 such answers the business was named 90.2% (751 of 833, 95% CI 88% to 92%). Business by business, the rate fell between 70% and 100%.
Getting the facts right is a separate test. For the 5 businesses whose facts were graded, between 15% and 35% of the answers that named the business got at least one fact wrong, counted per business across every question type. I publish that as a range and not a pooled rate, because the audits graded facts a little differently.
The businesses' own records gave the assistants room to get it wrong. In the 4 shorter snapshot checks, I compared each business's own website, its Google profile and its listings fact by fact. 4 of 4 disagreed with themselves somewhere, on at least 6 facts each. An assistant that reads two versions of your hours has to pick one.
That's why I treat accuracy as the first fix. It reaches the buyer closest to a decision, and the listings and pages involved are ones a business can correct or claim. The free What AI says about you template walks through the same check.
Want me to run it for you? Get a free AI answer snapshot.
Describe the problem,
and the name disappears.
Problem questions are the cleanest test of whether AI brings a business a stranger. The buyer has a problem and no shortlist yet. On those, the business was named 2.6% (16 of 623, 95% CI 2% to 4%). 6 of 9 were never named once.
Asked by name, AI answers named the business; asked about a problem, they rarely did
Pooled per answer across 9 businesses. The table carries the counts and both intervals.
| Question type | Named | Answers | 95% interval, answers | 95% interval, businesses | Range across businesses | Never named |
|---|---|---|---|---|---|---|
| Asked by nameThe question names the business ("Is [name] any good?") | 90.2% | 751 of 833 | 88% to 92% | 83% to 98% | 70% to 100% | 0 of 9 |
| LocalA service plus a nearby place, no business named | 38.7% | 271 of 701 | 35% to 42% | 15% to 53% | 0% to 70% | 2 of 9 |
| ComparisonOne option against another | 33.0% | 135 of 409 | 29% to 38% | 14% to 53% | 0% to 80% | 1 of 9 |
| Best-of and who-to-hireDiscovery questions that name no business | 12.5% | 124 of 994 | 11% to 15% | 5% to 25% | 0% to 70% | 1 of 9 |
| Price and buyingWhat it costs, where to buy, who to book | 14.3% | 104 of 725 | 12% to 17% | 6% to 29% | 0% to 65% | 1 of 9 |
| ProblemThe buyer describes what is wrong and asks what to do | 2.6% | 16 of 623 | 2% to 4% | 0% to 5% | 0% to 15% | 6 of 9 |
| Every question without the nameAll five types above except asked by name | 18.8% | 650 of 3,452 | 18% to 20% | 12% to 27% | 5% to 60% | 0 of 9 |
5 of the audits also graded whether the answer recommended the business as an option to use, a higher bar than naming it. Across every question without the name, it did about 15% of the time (376 of 2,472, 95% CI 14% to 17%), with per-business rates between 0% and 25%. I don't publish the recommended rate on problem questions alone, because one business supplies most of that count, and a pooled figure would mostly describe that one business.
Two intervals, because they answer different questions. The answer-level interval is answer-to-answer noise: how much the pooled rate would move if the same businesses were asked again. The business-level interval is business-set variation: it resamples the 9 businesses themselves, so it says how much the rate could move with a different set of businesses like these. It's wider, and it's the one I'd quote. The range shows how far apart single businesses sit, rounded outward to the nearest 5 points so no endpoint is one business's exact figure.
The answers cite
other sites.
Directories, review sites, forums and social profiles turned up in the answers for almost every business. These are the domains cited for at least 5 of the 8 businesses with source data. Each business's own site is left out of the count.
Businesses whose AI answers cited each domain, of 8
A domain appears only if it was cited for at least 5 businesses. No directory specific to one field reaches that bar, which is why none appear.
Reviews
In the 4 snapshot checks, I counted each business's Google reviews and the reviews of the competitors the AI answers named instead. 4 of 4 had fewer reviews than the median named competitor. That isn't proof that reviews cause naming, and with 4 businesses it can't be. It is consistent with what the answers cite.
Every crawler got in: 530 of 530.
I fetched each site's home page and money pages with the user agents of the AI and search crawlers: OpenAI's, Perplexity's, Anthropic's, Google's, Bing's, Apple's, Meta's and Common Crawl's. All 530 fetches across 9 sites returned HTTP 200. Blocked crawlers weren't why these businesses went unnamed.
One caveat. The fetches sent the crawlers' user-agent strings from a residential IP, not from the crawlers' own networks. A firewall rule keyed to verified bot IP ranges could behave differently. I check access once per site and then move on, as the first six audits suggested.
Fix what AI says,
then earn the mention.
- Ask the assistants about your business by name, ten questions, and write down every wrong fact. The free template has the questions and a log to fill in.
- Find where each wrong fact came from. In these checks it was usually a listing or profile, or two of the business's own pages disagreeing.
- Make one version of each fact true everywhere: website, Google profile, Bing Places, Apple Business Connect, and the directories that answers in your field cite.
- Ask real customers for reviews on the platform the answers cite, and reply to them. Don't post fake reviews or tie a reward to a positive one; the FTC's consumer review rule (16 CFR Part 465) covers both (checked October 27, 2026).
- Re-ask the same questions later, the same way. Answers change run to run, so compare counts, not single answers.
Quotable findings to lift as written.
Each line carries its count or range, so it survives being quoted on its own.
- Across 833 AI answers to questions that named the business, 9 businesses were named 90.2% of the time (751 of 833), per my Q4 2026 AI answers field report.
- On problem questions that named no business, AI answers named the business 2.6% of the time (16 of 623, 95% CI 2% to 4%), and 6 of 9 businesses were never named.
- For the 5 businesses whose facts were graded, between 15% and 35% of the AI answers that named a business got at least one fact about it wrong (per-business range, rounded outward).
- 4 of 4 businesses I checked had facts that disagreed across their own website and listings, at least 6 conflicting facts each.
- 4 of 4 businesses had fewer Google reviews than the median competitor the AI answers named instead.
- All 530 fetches with AI and search crawler user agents, sent from a residential IP, across 9 business sites returned HTTP 200.
How the answers were graded, and the limits.
The pool
- 9 businesses, each with its own locked panel of roughly 40 to 60 buyer questions. The full audits built their panels from the client's own Google Ads search terms and Search Console queries, plus customer phrasing. The snapshot panels came from public sources only: search autocomplete, the business's own website, and a few fact-check questions every snapshot asks. Some were full audits for clients, some were free snapshot checks.
- Baseline runs only. Re-tests after fixes aren't pooled, because a re-test asks whether one business moved. My own 30-day re-test is in its own post.
- My own site is not in the pool. Its numbers are mine, so I report them in full in the six-audit post.
- No industry rates. I don't say which fields are in the pool, and nothing here is a rate for any field. Only finished baselines are pooled.
Engines and runs
- Assistants: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews, Microsoft Copilot, Perplexity. Not every assistant ran in every audit. Quotas, CAPTCHAs I didn't solve and sign-in walls got in the way, and the shorter snapshot checks ran fewer engines.
- Runs: up to three runs of each question on each assistant, because the same question asked twice gets different answers. Quotas cut some engines short, and each audit records where.
- Sessions: clean sessions where possible. Where an audit found an assistant leaking account context into its answers, those answers are out of the pool. Answers with no response are out too.
- Per-assistant rates aren't published. The engine mix differs between audits, so a per-assistant figure would mostly reflect which businesses that assistant was run for.
What was graded
- Named: the answer names the business anywhere in its text.
- Recommended: the answer presents the business as an option to use. 5 of the 9 audits graded this.
- Wrong fact: at least one statement about the business that contradicts its live website or verified profile. Facts were graded for 5 businesses.
- Question types: asked by name (the question names the business), problem, best-of and who-to-hire, comparison, price and local questions. "Every question without the name" is all but the first.
- Conflicting facts: one row per fact on which the business's own site and listings disagree, counted in the 4 snapshot checks.
How rates were pooled
Each rate is the qualifying answers that named the business, summed across businesses, over all qualifying answers. Every rate carries two 95% intervals: a Wilson interval over answers, which stays honest near 0% and 100%, and a bootstrap that resamples whole businesses 10,000 times, because answers about one business aren't independent of each other. Beside each rate is the per-business range, rounded outward to 5-point steps, and the number of businesses at zero.
A script builds every number on this page from the raw audit files and refuses to output any figure that rests on fewer than 3 businesses. It also withholds a pooled count when one business supplies most of it, and it never outputs a figure for a group of businesses. Industry directories are left off the domain chart for the same reason.
Limits
- 9 businesses is small. The answer-level interval covers answer-to-answer noise only. The business-level interval covers variation from one set of businesses to another like it, but neither says how well 9 businesses stand for every business.
- The businesses chose to be audited, or asked for a snapshot. Owners who ask are more likely to suspect a problem.
- Engine setups differed between audits, and assistants change week to week. Treat the numbers as a dated reading, which is why each edition carries a date.
- Location matters. Some assistants answered for the tester's location rather than the business's market. Where the audit could, Google's surfaces were pinned to the business's city. Other answers recorded the tester's location and were flagged.
- Reviews and conflicting facts come from 4 businesses. They show a pattern worth checking, not a measured effect.
Editions are dated and refreshed.
I refresh this report as new audits come in, and a new edition ships only once at least three new businesses have joined the pool. The changelog gives the pool size as a band, and I don't publish what changed between editions business by business. Both rules stop anyone working out one business's numbers by subtracting one edition from the next.
- Q4 2026 (): 10 to 15 businesses. First edition.
- Q4 2026 (): Correction: one business that was audited without asking for it was removed; all figures recomputed.
This edition replaces the figures in my October 4 post on six audits. That post used every question without the name as its headline and included my own site; the note at the top of the post explains the difference.
How to cite this report.
Conner Crowe, "AI Answers Field Report," Q4 2026 edition, connercrowe.com, October 27, 2026. https://connercrowe.com/benchmarks/ai-answers-field-report/
Please link to this page so readers can check the counts and the edition. The CSV has every pooled figure here with its definition, count, interval and per-business range. It has no per-business rows.
Data licensed CC BY 4.0, attribution to Conner Crowe with a link.
What owners ask.
How often does ChatGPT name a business when people ask who to hire?
Does AI get business information wrong?
Is blocking AI crawlers why my business is not in AI answers?
Why are these numbers different from your October 4 post?
More benchmarks
- Ads inside Google AI Mode answers How often Google's AI Mode puts a paid ad inside a local hiring answer.
- Google Ads account benchmarks Cost per click, cost per lead and account health across live accounts.
- All benchmarks The index, and the rules every report is held to.
See what AI tells
buyers about you.
Send your website and your market. I'll ask 10 questions your buyers ask on ChatGPT, Google AI Overviews and Perplexity, and within 48 hours you'll have what each one says about you, which facts are wrong and where they come from. Free, written by me, no call needed.