We Checked 284 Google Maps Results. 1 in 4 Was Not the Business We Searched For.

How we measured it

We took 12 English-language keyword searches already in our corpus and sampled up to 30 random businesses from each, one per distinct domain so that a chain with forty branches could not dominate a keyword. For every business we fetched the homepage, stripped it to visible text, and judged it against the keyword with a four-way rubric. Every verdict required a verbatim quote from the page as evidence, so no judgement rests on a business name alone.

• Match — the business primarily is the thing searched for.
• Adjacent — a related trade that is not it. Roof cleaning for "roofing contractor". A supplier to the trade rather than someone practising it. A generalist listing the service among twenty others.
• Mismatch — a different business entirely, or a directory or aggregator rather than an operating company.
• Unclear — the page never says what the business does.

The rubric explicitly refuses to downgrade a business for being small, for calling itself a studio or a practice rather than an agency, or for specialising in one industry. An earlier version did all three, and it understated the match rate, so we fixed it and re-ran.

Sites we could not read were recorded separately and excluded from the denominator, because that is a fact about the website, not about whether the business matched. It is a large fact on its own: 76 of 360 sampled sites, 21.1 percent, were dead, blocked or empty, which is consistent with our earlier website churn study.

Finding one: the keyword decides everything

A precise trade name lands. A broad or ambiguous one falls apart. "Home remodeling contractors" returned house cleaning services, a hotel, a pest control company, a mold remediation firm, a commercial real estate brokerage and an apartment complex. "Staffing agency" returned two theatres, a car rental branch, a UPS Store and a Marine Corps recruiting office.

Finding two: the obvious fix works better than we expected

We built this study expecting to show that the fields cannot save you, and that only reading the website can. That is not what the data said, and it would have been easy to leave this section out.

Every Google Maps result carries a business category. Filter the results down to the categories that the genuine businesses actually carried, and almost all of the junk disappears: only about 4 percent of what survives is still the wrong kind of business. A theatre does not carry the category "Employment agency". If you are building a list and you are not filtering on category, start.

Finding three: that fix quietly deletes real businesses

It is not free, and the cost is invisible, which is the worst kind. The same category filter also discarded 24 of the 205 genuine businesses, 11.7 percent. We read all 24 by hand. They are real, and several are exactly the customer you were looking for:

• Big Brand Tire & Service, whose own page says "Tire Shop & Auto Repair in Phoenix, AZ", is categorised Tire shop. Dropped from an auto repair list.
• J & J Remodeling Team LLC, "a locally owned remodeling company", is categorised Remodeler. Dropped from a home remodeling list, because the category that dominated the genuine results was Home builder.
• RSU Contractors, "expert kitchen remodeling, bathroom renovations, and home additions", is categorised General contractor.
• A pediatric dentist, an orthodontist and a periodontist all dropped from a dentist list, because Google files each specialism as its own category.
• Econo Lube N' Tune & Brakes, doing "automotive repair and maintenance work", is categorised Oil change service.
• A plumber categorised Home help. Nobody would have guessed that one in advance.

This is the trap: the category is one self-chosen label from a list of thousands, and it frequently describes a specialism rather than a trade. Dentist, Orthodontist, Pediatric dentist and Dental implants periodontist are four separate Google categories for what a dental supplier would call one market. To filter without losses you would need to know the whole family of categories in advance, for every trade you sell into. Some of that family is guessable. Home help for a plumber is not.

So both things are true at once, and only saying the first one would be a sales pitch rather than a finding. Category filtering is a good, cheap, underused tool that removes most irrelevant results. It also silently removes about one in nine of the businesses you wanted, and you will never see the ones it took.

Finding four: we could not compare the two methods, so we ran a second study

Everything above had a hole in it. The verdicts in this study were produced by one language model reading each homepage, which means that model both wrote the answer key and is one of the methods being graded against it. Scoring the category filter's mistakes against those verdicts and declaring the model the winner is circular, and we were not willing to publish it that way.

So we built a referee. The same 284 pages were re-fetched and re-judged by a different model (GPT-4o rather than GPT-4o-mini), asked a different question (a single yes or no, from a salesperson's point of view, with no rubric and no mention of categories), and blind to the first verdict. 267 pages could be read by both. Then we scored both methods against the referee.

First, do the two raters agree at all? 89.5 percent raw agreement, Cohen's kappa 0.765, which is substantial agreement rather than either near-identity or noise. Without that, none of the rest would mean anything.

*The category filter needs a set of acceptable categories. We built that set two ways: from the first rater's verdicts, which is what you could actually build in practice, and from the referee's own answers, which leaks information in the category filter's favour and is the best score it could ever get. It loses both ways.

Precision is a tie. Recall is not close. Of the businesses the referee said were genuinely what was searched for, filtering on category kept 87.4 percent and reading the website kept 98.8 percent. In counts: the category filter missed 21 real businesses, reading the site missed 2.

That 12.6 percent miss rate is worth pausing on, because it was produced by a completely separate run and it lands on top of the 11.7 percent we measured earlier and checked by hand. Two independent routes to the same number is the strongest thing in this post.

So the claim we would not make before, we will make now: reading the website does not catch more junk than a category filter, it keeps the businesses the category filter throws away. Roughly one in eight of them, silently.

Where this comparison is still weak

One honest limitation, and it is the one that matters. Both raters are language models reading the same page text. They share a method, and two instances of the same method will naturally agree with each other more than either agrees with a category field. Some of the gap above is that shared bias rather than genuine superiority, and we cannot say how much.

What would settle it properly is a panel of human raters scoring the same pages. That is a slower and more expensive study than a two-dollar model run, and we have not done it. Until we do, treat the direction as well supported and the exact size of the gap as provisional.

The disagreements between the two raters are also worth naming, because they are not random errors. They cluster almost entirely on one boundary: full-service digital agencies returned for "social media marketing agency". The first rater counted them, the referee did not. That is a definitional disagreement about how broad a category word is, not a factual mistake by either. If you prospect agencies, that boundary is yours to draw, and it is exactly the kind of judgement no category field can make for you.

What to actually do with this

Three things follow, none of which require our product.

Use precise trade keywords, not category words. "Roofing contractor" is a trade. "Home remodeling contractors" is a category of work that dozens of unrelated trades touch, and it returned a list that was 72 percent noise. If you must run a broad keyword, expect to qualify by hand.

Filter on category, then look at what you dropped. The filter is worth running. Pull the excluded rows into their own view instead of deleting them, and you will find a tire chain and three dental specialists sitting in it.

Judge the business by its website, not by its label. Every claim in this study came from a verbatim sentence on the company's own homepage, because that is the only place a business says what it actually does. It is also why we enrich every result from its live website rather than trusting a category field, and why every field we ship records how we know it.

Methodology notes for anyone citing this: 12 English-language Google Maps keyword searches, up to 30 US businesses sampled per keyword, deduplicated to one row per registrable domain, 360 sampled and 284 judged; homepages fetched in a single pass on August 9, 2026 with a plain-HTTP fetcher; judgements by gpt-4o-mini at temperature 0 with a four-way rubric requiring a verbatim quote as evidence; 76 unreachable sites excluded from the denominator and reported separately; all 75 non-match verdicts and all 24 category-filter losses read by hand. Errors were found in both directions and roughly cancel, so treat the headline as roughly 7 in 10, plus or minus 3 points, not as 72.2 percent to one decimal.

Second rater: the same pages re-fetched and re-judged the same day by gpt-4o at temperature 0, asked a single yes or no question under a differently worded prompt that mentions neither the rubric nor the categories, and blind to the first verdict; 267 pages readable by both raters; raw agreement 89.5 percent, Cohen's kappa 0.765. Precision and recall for both methods computed against the second rater, with the category filter's acceptable-category set built both from the first rater's verdicts and from the second rater's own answers, the latter being a deliberate leak in the category filter's favour. Both raters are language models reading the same text and therefore share a method bias of unknown size; a human panel would be the stronger test and has not been run. Cite freely with a link.