Rainmaker.
Compare and Alternatives · 4 min read

The lookalike bug: why AI GTM tools send you to the wrong buyers

When your ICP engine returns competitors as "ideal customers," the problem isn't the data. It's the targeting model.

There is a bug that ships in most AI GTM tools. It is not a data quality problem, though bad data makes it worse. It is a targeting model problem, and it has a specific name: the lookalike bug.

Here is what it looks like in practice: you paste your company URL into an AI ICP tool. It analyzes your existing customers (or your description of them), runs an embedding similarity search across a company database, and returns a list of "ideal customers." Those companies look like your customers - same industry, similar size, similar tech stack. Some of them are your competitors.

We watched this happen firsthand during our testing of Explee (AutoGTM). We ran our own company through their ICP engine. The results included companies that were our direct competitors. It also returned a nuclear energy facility tagged as a real-estate company.

This is not a data error. It is the expected output of lookalike targeting applied to the wrong question.

How lookalike targeting works (and why it sounds right)

Lookalike targeting starts from a sensible premise: if you know who your customers are, you can find more companies like them.

The technical implementation uses embedding models. You take your customer list, generate vector representations of those companies (industry, size, growth signals, tech stack, job titles), and run a similarity search across a large company database. The output is a ranked list of companies that are similar to your customers.

This is exactly how Meta's lookalike audiences work, which is why the concept feels familiar. For paid social advertising, it works reasonably well. For B2B outbound, it has a structural problem.

The problem: lookalikes find companies that resemble customers, not companies ready to buy

The similarity search answers the question: "Which companies look like our customers?"

It does not answer: "Which companies have the specific pain our product solves, right now?"

These are different questions. A company that looks like your customer may:

  • Already use a competitor (making them actively hostile to your outreach)

  • Be the right size and industry but not have the specific trigger that creates urgency

  • Be your direct competitor (who has the same profile as your best customers, by definition)

  • Have solved the problem themselves internally

  • Be in a market where your pricing does not fit

Lookalike targeting is optimized for demographic resemblance. Buying behavior is driven by situation, not demographics. The company that urgently needs what you sell today is often not the one that looks most like your best customer - it is the one that just had a triggering event your best customer had before they bought from you.

What we observed: the ICP engine returns competitors

We ran our own company URL through Explee's ICP search during our competitive research phase.

The results included companies in our space - direct competitors building AI SDR tools. They look like us on the surface: similar size, similar industry category, similar audience. A pure similarity model has no way to exclude them. From the data's perspective, they are ideal matches.

We also saw a nuclear energy facility categorized as a real-estate company in the database. This points to a second problem: the underlying data quality in large scraped databases, even at scale (Explee's database covers 105M companies by their documentation), is uneven. Misclassified companies get through similarity searches because the model does not know the tag is wrong.

Both outputs - competitors as buyers, misclassified companies - are predictable consequences of the lookalike approach. They are not bugs in implementation. They are bugs in the method.

How BuyerPersona is different

Rainmaker does not use embedding similarity to find leads.

Instead, it builds a BuyerPersona: a structured model of who actually buys from you, defined in terms of pains, purchase triggers, and disqualifiers. Not "companies that look like your customers" but "companies that have the problem you solve and a specific reason to solve it now."

The difference in what gets structured:

Lookalike targetingBuyerPersona
Industry, size, tech stack, growth signalsSpecific pains the buyer experiences
Similarity to existing customersPurchase triggers (events that create urgency)
Static profile matchActive disqualifiers (who will never buy, filtered out upfront)
Who looks like a buyerWho is a buyer right now

When a lead is sourced against a BuyerPersona, Rainmaker shows you the reasoning: why this company fits the persona, which trigger was detected, what disqualifiers were checked. This is shown for free, before you approve the batch.

If the reasoning looks wrong, you reject the lead and the persona gets tighter. This feedback loop does not exist in lookalike targeting, because there is no reasoning to correct - just a similarity score.

Why this matters in practice

Wrong buyers do not just waste time. They actively degrade your outbound program.

Cold email sent to companies that are similar-but-wrong burns in two ways:

  1. Low reply rates (the companies are not in-market)

Spam complaints (recipients who are not interested, or who recognize the approach as a mass blast, mark it as spam) Spam complaints damage your sending domain reputation. Once you cross a threshold - typically 0.1% complaint rate for Gmail - deliverability drops across your entire list, including the buyers who would have replied.

Getting targeting right is not just about conversion. It is about protecting the infrastructure your whole outbound program depends on.

Rainmaker models who buys so that outbound only goes to companies where there is a real basis for the conversation. That is what keeps the list small, the targeting tight, and the domain healthy.

Questions this post answers

In short.

What is the difference between a BuyerPersona and a lookalike model?

A lookalike model finds companies that resemble your existing customers using embedding similarity - similar industry, size, tech stack. A BuyerPersona is a structured model of who actually buys: their specific pains, the events that trigger a purchase, and the characteristics that disqualify a company. Lookalike targets who-looks-like-a-buyer. BuyerPersona targets who-is-a-buyer-now. The practical difference is that lookalike returns similar companies (including competitors); BuyerPersona returns companies with active purchase reasons.

What is "lookalike targeting" in AI outbound tools?

Lookalike targeting generates vector representations of your existing customers and runs a similarity search across a company database to find companies that match the profile. It is borrowed from paid social advertising (Meta lookalike audiences) and applied to B2B outbound. The structural limitation is that demographic similarity does not predict buying intent - a company that looks like your customers may be a direct competitor, may already use a competing solution, or may not have the specific problem your product solves. 1. 2.

Does Rainmaker use any lookalike matching?

No. Rainmaker builds a BuyerPersona from your company URL and a discovery conversation, then sources and scores leads against that persona. Each lead gets a per-lead score and reasoning (which pains match, which triggers were detected, which disqualifiers were checked). This reasoning is visible to you before you approve the batch, at no extra cost.

How does Rainmaker explain why a lead fits?

For every lead in your approval batch, Rainmaker shows the score and the reasoning in plain language: which pain from the BuyerPersona this company has evidence for, what trigger event was detected, and which disqualifiers were checked. You can reject any lead, and rejected leads with reasons make the persona sharper over time.

See the receipts on your own domain.

Paste your URL and read the persona and the sourced drafts it writes for your market. No card for the 14-day trial.

You might also read