Articles & Case Studies
Beyond Lookalikes: ICP Fit and Full TAM Search
Lookalike search retrieves similar companies, which can leave gaps when mapping every account that meets an ICP. DiscoLike is planning model training that uses ICP validation results to make fit part of retrieval.
Lookalikes are no longer the state of the art for account discovery. Agents are asking for the full total addressable market (TAM): every company that fits an ideal customer profile (ICP).
That request changes what the search needs to accomplish. Returning companies similar to a customer or description is useful. Mapping the qualifying market also means finding businesses that fit the criteria despite looking different from the examples.
How account discovery moved from categories to lookalikes
Targeting moved from broad industry classifications such as NAICS and SIC codes to keywords, shaped heavily by LinkedIn search and company descriptions, then to lookalikes. Each step let us get more specific about the businesses we wanted to reach.
The vectors and models behind lookalike search are optimized for similarity: find companies like this one, or companies that look like this description. That is a powerful way to discover candidates.
Agents now need the full set of companies that meet the actual criteria. Similarity alone cannot reliably establish that set.
Why ICP validation cannot close every coverage gap
Checking each result for ICP fit improves the list you already have. You can remove companies that fail the criteria and keep those that qualify.
But you can only validate the companies your search found. Qualified companies left out of the results remain invisible to that validation step.
A skilled human can recognize a gap, change the description, and add searches to cover another part of the market. An agent does not reliably recognize and repair those gaps on its own. When the job is full TAM mapping, those omissions matter as much as the obvious mismatches.
Three ICP requests that go beyond similarity
Consider these requests:
- Consumer packaged goods (CPG) companies operating in five or more countries and managing five or more brands.
- Manufacturers with large open factory floors and critical security needs.
- Large businesses within 20 miles of any private airport on a given list.
These combine business activity with operating scale, physical characteristics, or geographic relationships. A company can meet the conditions without closely resembling a particular seed company.
Lookalikes can help discover candidates. Finding the qualifying market requires the search to account for the criteria themselves.
Training a search model for ICP fit
We have accumulated hundreds of thousands of ICP requests and millions of records with yes-or-no ICP fit validation results. That gives us training material for a model focused specifically on fit, including distinctions that similarity search misses.
The goal is to make ICP fit part of the search, bringing back more of the qualifying market from the start. Today, teams can sift through several times the records they need, spend research budgets distilling the subset that fits, and still leave gaps.
DiscoLike is preparing to commit a sizable R&D budget to GPUs and new model training. This is the research direction we are pursuing; the broader retrieval capability is still ahead of us.
Agents are coming, ready or not. They need discovery that can do more of the market-mapping work before the validation bill starts growing.
Have an ICP that exposes the limits of lookalikes? Bring it to a DiscoLike demo and help us test where account discovery needs to go next.
Related posts:
