You can build a TAM list for a niche that directories miss from Google results: we bought them through Serper, crawled each homepage and had a model judge fit, which gave nearly 5,000 qualified accounts from one September 2026 sweep. Query wording decided most of the hit rate, and the judge only worked after we fixed both the model and the rubric.
The question here is whether you can build a usable total addressable market list for a niche market out of plain Google results, and what it takes to trust the list once you have it. We tried it in September 2026 for a market that doesn't line up with industry codes, buying the search results through Serper's API.
Directories had copied each other
We started where most people start: six "top vendors in X" directories and listicles for one segment. Together they named 43 companies. None of the handful of existing customers we had in that segment appeared on any of the six.
The lists overlapped so heavily that a capture-recapture estimate of the population they were drawing from came out at about 15 companies, smaller than the 43 we had already seen. That only happens when sources copy the same well-known names. A single query written the way an operator in that segment describes their own work found 8 companies that were on none of the lists.
That gave us the acceptance test we used for everything after: does the method rediscover customers we already have?
Phrase the query from the target's side
The first full round used 20 hand-written categories and 468 queries, cost under $1 of Serper credits, and found all but one of the customers the directories missed. That was enough to build out a full frame. We had a model read the websites of about 100 customers and propose the categories they fell into, then wrote queries for each.
Hit rates by category ranged from 95% down to 6%, and the wording of the query explained most of that spread. The worst query was a topic phrase, and Google answered it with media sites and blogs. Next worst were queries written from the customer's side of a service, which brought back the businesses that buy the service instead of the companies that provide it. As an illustration from outside our market, a search for "office catering" written the way a buyer would phrase it would likely surface offices rather than caterers. Writing each query in the words the target company uses on its own homepage fixed most of it.
One category surprised us. The segment where we had won the most deals came in at 56%, near the bottom of the useful range, which is why we derived categories from what customers' sites say rather than from where we thought we were strong.
Crawl the homepages yourself
Search results give you a domain and a snippet, which isn't enough to judge fit. Our own crawler fetched the homepages and got usable text from 79% of them.
An HTTP 403 usually meant a site blocked bots, and those were often real operating companies, so we sent them to review instead of dropping them. We only dropped a company on evidence that it was gone, such as a DNS failure, and a refused page doesn't count as that evidence. For blocked sites, Serper's own result snippets made a workable fallback description at very low cost.
The judge needed two fixes
The first judge was a small model, Haiku. It returned the same templated answer for 150 different companies and rated nearly all of them qualified. We threw that run away and moved judging to Sonnet, checking each batch for how many distinct answers it produced.
The second problem was ours. An earlier fit rubric, given 23 companies we knew were good fits, rejected all 23. The rejections opened with some version of "while they handle X", because the rubric asked whether a company called itself the thing we were looking for, and our customers don't. After the rewrite it passed 18 of the 23.
We then calibrated the new rubric on accounts we already knew. Customers and open deals were judged to have a clear need about twice as often as random companies, which was the lift we needed to rank on it. Two runs of the same judge on the same companies still disagreed about a fifth of the time, so we treat one verdict as a sort order and re-judge anything near the line.
What the sweep produced
The judged sweep ran in September: about 9,000 queries, about 12,000 unique domains, about 8,000 companies judged and nearly 5,000 qualified. Nearly nine in ten of the qualified accounts were not in our CRM. "Qualified" means the model judge passed them after the calibration above; we didn't hand-check them all.
The CRM check mattered. Of the qualified accounts we matched against HubSpot by domain, about one in ten were already there, and far fewer had ever been contacted.
What it cost
Search cost about $9, at roughly $1 per 1,000 Serper queries. The crawler was our own script. Model API calls to read pages and judge fit came to about $60 to $90, and some judging ran on a flat-rate subscription on top of that. Search plus model calls works out to roughly 1.4 to 2.1 cents per qualified account. That figure leaves out the developer time to build and fix the pipeline, which was the biggest cost by far.
Where lookalike tools fit
Before this, we tried lookalike search from Ocean.io on the same market, and few of its results passed our fit rule. That test, and why our niche was a hard case for it, is on the Ocean.io review. The two numbers came from different graders (a written rule checked against each company's site there, a calibrated model judge here), so don't set one rate against the other.
Who it's for
| Build it this way if | Buy a list instead if |
|---|---|
| Your market is a niche that directories cover badly | Your buyers sit neatly in standard industry codes |
| Someone can run Python scripts and read JSON | Nobody on the team will own a pipeline |
| You can describe what a target company says on its homepage | You can only describe the market by firmographics |
Run it yourself in 30 days
- Pick one segment and check whether directories list your own customers in it.
- Write 20 queries in the words a target company uses about itself.
- Pull results from Serper, dedupe by domain and crawl the homepages.
- Hand-check 50 domains before any model judges them, so you know the real hit rate.
- Write the fit rubric and test it on 20 customers. If it rejects most of them, fix the rubric first.
- Run the judge twice on a sample and measure how often it disagrees with itself.
- Match against the CRM by domain, including www and subsidiary domains, before anyone sends an email.
Once the accounts exist, cost per verified contact covers finding people at them, and hiring signals from job posts covers deciding which to contact first.
Questions buyers ask
How much does Serper cost for building a TAM list?
Search was the cheap part: about $9 for some 9,000 queries at roughly $1 per 1,000 in September 2026. Model calls to judge fit cost $60 to $90 more, and building the pipeline cost far more than either.
Can you build a TAM list from Google search results?
Yes, if you write queries in the target company's own words and crawl the homepages before judging. Our sweep gave nearly 5,000 model-qualified accounts, and query wording moved category hit rates between 95% and 6%.
Why don't industry directories work for niche markets?
In our test they had copied each other. Six directories named 43 companies, none of them our handful of customers in that segment, and the overlap implied a source population of about 15.
Can a model judge account fit reliably?
Only after testing it on companies you already know. Our first rubric rejected 23 of 23 known good fits, a small model gave templated answers, and two runs of the final judge disagreed about a fifth of the time.
Do you need a GTM engineer to run this?
You need someone who can run a few Python scripts: one to call Serper, one to crawl homepages, one to send pages to a model with a rubric, and one to match domains against the CRM.