By Zach Wright, Cofounder at Syft AI
A lookalike model answers a reasonable question. Your closed-won accounts share attributes, so find more companies with those attributes and you should find more revenue. That approach is well understood and the tooling is mature, which is exactly why the distinction between lookalike similarity and what Syft AI calls a value match, a company showing current evidence of the problem you solve, gets collapsed so often in practice. Upload 50 to 100 of your best customers, a lookalike model extracts shared traits across firmographics, technographics, hiring velocity, and funding stage, then scores your addressable market by resemblance and hands back a ranked list.
What comes back is a population statement. These companies resemble the population that bought from you. That is genuinely useful for territory design, for media targeting, and for arguing that a segment deserves headcount. It is a different thing from knowing that a specific company has a problem you can solve this month, and treating the two as interchangeable is where a lot of outbound turns into volume.
Every closed-won account in your seed list carries two categories of information.
The first is durable. Industry, employee count, revenue band, headquarters, tech stack, funding history. These attributes were true before the deal, they were true during the deal, and they're still true today. They persist in your CRM because they persist in the world.
The second is transient. A controller inherited three sets of books after an acquisition. A compliance deadline landed with a real penalty attached. A VP arrived with a mandate and eight months to show something. The head of data quit and took the pipeline knowledge with them. These conditions existed for a window, they caused someone to go looking for a solution, and then they resolved. Your CRM records the deal, not the condition.
A similarity model can only learn from what survives in the record. It preserves the durable attributes with high fidelity and discards the transient conditions almost entirely, because by the time the model runs, the transient conditions have already stopped being observable in your data. The result is a model that faithfully reproduces the part of the pattern that did not cause the purchase.
This is not a flaw in anyone's implementation. It's a property of training on outcomes recorded as static records. The same limitation applies to embedding-based similarity, which measures semantic distance between company descriptions rather than trait overlap. A vector that captures what a company is will not capture what is happening inside it, because company descriptions are written to be stable.
There's a second-order problem stacked on top. Your seed list is composed of accounts you closed, which is a subset of accounts you targeted, which is a subset of the market you decided to look at. A lookalike model trained on that set learns your historical targeting decisions along with your win pattern, and it can't distinguish between them. Run it repeatedly and it narrows toward the segments you were already covering. The accounts absent from both your outbound lists and your inbound flow stay absent.
Even a perfectly specified similarity model faces a base rate that most account scoring quietly ignores. B2B categories turn over on multi-year cycles. If companies replace a given service roughly every five years, only about 20 percent are in market in a given year and something closer to 5 percent in a given quarter. John Dawes at the Ehrenberg-Bass Institute has been consistent that this ratio is a heuristic rather than a constant, and that you should recalculate it against your own category's interpurchase time.
Behavioral data points the same direction. Analysis across 19 million monitored accounts found roughly 40 percent showing some in-market activity, only 2 to 4 percent genuinely ready to purchase, and 60 percent showing no buying signals at all.
Apply that to a lookalike output. A list of 1,800 accounts at 90 percent similarity contains roughly the same proportion of companies with an active reason to talk to you as any other slice of your addressable market.
Similarity ranking does not change the base rate. It changes which cold accounts you work first.
These are anonymized composites drawn from patterns we see repeatedly in value match evidence.
The pattern in all three is that the differentiating variable is a condition, not an attribute, and conditions live outside your CRM.
The argument above is testable in about a week, and the test is more persuasive to a sales leader than any vendor claim. Hold the seller, the territory, and the messaging constant. Vary only how accounts were selected.
Take 100 accounts selected by top similarity score and 100 accounts selected because they show current public evidence of an active problem you solve. Track five numbers across both cohorts over 30 days:
Second meetings are the number that matters. Reply rate measures whether your subject line worked. A second meeting measures whether the buyer thought the first one was worth their time, which is the real confirmation that the account had the problem you thought it had.
The alternative starts from a different input. Rather than learning from the attributes of accounts you closed, the system learns what you sell: the specific use cases, the value propositions, and the win stories that usually live with your top two reps. That profile becomes the criteria.
The evaluation then runs against current public evidence rather than static company records:
The output is an account plus the reason, with the source and the date behind it. That reason is the part a similarity score cannot produce, and it's also the part a seller needs in order to write a first line that earns a reply.
One practical consequence: this approach doesn't require a large seed set. Lookalike modeling generally needs 30 to 50 closed-won accounts before the patterns are statistically meaningful, which puts it out of reach for a new product line, a new segment, or a company selling into a market it hasn't won in yet. Evaluating evidence against what you solve works from day one, because the criteria come from your value proposition rather than your history.
Syft is a context layer for B2B sales teams. It ingests what your company sells, then evaluates public evidence against that profile and returns value matches: companies actively working on the problem your seller solves, each with the evidence behind it, source URLs and dates, the reason the account fits that seller's specific context, and which value proposition applies.
Because detection is based on what is happening at the company rather than on resemblance to your existing customers, the output includes accounts that no similarity model would surface. A company can be an outlier on every firmographic dimension and still be the best conversation you'll have this quarter if the problem is live and someone owns it.
Sellers work these accounts in the Syft app, which refreshes weekly by seller or territory. Teams running agent-based outbound consume the same records through the Syft MCP or the Value Match API, which gives a sequencer or AI sales agent the account, the reason, and the evidence in one pass.
Lookalike modeling and value matching are answering different questions. One tells you where your market probably is. The other tells you which specific companies have a reason to take the meeting right now. The first is a planning input. The second is a working list.
What is the difference between a value match and a lookalike account?
A lookalike account resembles your existing customers on durable attributes such as industry, size, and tech stack. A value match is a company showing current evidence of the problem your product solves, with the source and date attached. Resemblance is a statement about the population. A value match is a statement about that specific company this week.
Are lookalike models useless for B2B?
No. They're effective for market sizing, territory planning, segment prioritization, and paid media targeting, where population-level accuracy is what you need. They break down when treated as a prospecting queue, because resemblance does not indicate that anything is happening inside the account.
Why don't lookalike companies match your ICP?
They usually do match the ICP as written. The issue is that most ICP definitions describe attributes rather than conditions, so an account can satisfy every ICP criterion while having no active reason to buy. Adding contextual fit and timing to a scoring model helps, though those dimensions are typically the smallest weights in the rubric and the fastest to go stale.
Can't you just layer intent data on top of a lookalike list?
That helps with timing and it's the standard recommendation. Topic-level intent tells you someone at an account consumed related content, which usually means an evaluation is already underway and requirements are forming or formed. It gives you a score without telling you what the account is dealing with, which leaves the seller to guess at the reason. Evidence of the operational situation appears earlier and comes with the reason included.
How much closed-won data do you need for this to work?
None. Similarity and propensity models generally need 30 to 50 closed-won accounts before patterns are reliable. Evaluating public evidence against your value propositions works for a brand new product or a segment you've never sold into, because the criteria come from what you solve rather than from who you've already closed.
How do you verify that a value match is real?
Every match carries its evidence: the source URL, the publication date, and the specific language that triggered it. A seller can read the job posting or the earnings transcript before deciding to work the account, which is a check that a similarity score does not offer.
