How to Tell If a Buying Signal Is Real

We record what a business's website and Google listing look like every time we check them, and we compare each check with the one before. When something differs, that difference becomes a signal: added a CRM, switched chat widget, review surge, rating dropped. A salesperson uses it as the reason to write today.

That use is exactly why a wrong signal costs more than a missing one. If you open with "saw you just switched your email provider" and they didn't, you've told the prospect your data is wrong before you've told them anything else.

1. Did the business change, or did the detector?

Every tool that reads websites improves its detection over time, adding new booking systems, new chat widgets and new CRMs. The day a tool learns to recognise a vendor, every business already using it looks like it "just added" it.

So the first question for any change is whether both observations came from the same version of the detection code. Ours stamps every observation with a pipeline version, and a change that spans two versions is never shown as a signal. When our detection gets better, you get better data on your leads, and not a wave of "adoptions" that never happened.

2. Is the "before" value real?

A change has two ends, and people only look at the new one.

In raw data, a failed read can make a property management company appear to go from 0 Google reviews to 293 in sixteen seconds. It didn't, of course. Google occasionally returns a listing without its review block, so for one read the count comes back as zero with no rating, and a moment later the real number arrives.

The check is simple once you know to make it. A listing with no rating and zero reviews tells you nothing about reviews, so we don't record it. On the website side, the same principle applies to a page that failed to load: a failed read isn't an absence, and it must never become the "before" of a removal.

3. Is the change big for this business?

A rule of "ten new reviews" sounds reasonable until you point it at a shopping mall. Going from 14,048 Google reviews to 14,073 clears it, and that's 0.18%. Malls, museums, hotels and music venues clear a fixed count just by being busy.

A change has to be measured against the business it happened to, and against the time it took. A roofer going from 31 to 53 reviews in 46 days is growing fast. The same 22 reviews over five months on a listing with 1,000 is noise. We count a review surge only when it's at least 5 new reviews and at least 5% growth per 30 days. The full rule and the worked examples are in what counts as a review surge.

4. Did the thing change, or only its name?

"Switched email provider" is one of the easiest signals to get wrong, because two very different things look like a switch.

One is a relabel. If one check stores a mail host as "simply.com" and the next as "Simply.com", a straight text comparison calls that a switch. The other is a spam filter: a business on Microsoft 365 puts Trend Micro's filtering in front of it, the MX record changes, and the company's email provider doesn't.

If a signal compares values, ask what it compares. Ours reduces both sides to the actual mail host first and ignores moves to or from a security gateway.

5. Does the label claim more than the evidence?

A newly detected ad pixel is easy to call "started running ads". A pixel shows that ad tracking was installed on the site. Plenty of businesses install one and never spend a cent, or stopped years ago and never removed it. Our signal is called "Added ad tracking", because that's what can actually be seen.

The same goes for a Stripe domain-verification record appearing in DNS, which is easy to call "preparing online payments". Existing Stripe users add those too, so ours is called "Verified domain with Stripe".

Read the evidence behind every signal name you're sold. If a vendor can't tell you what was observed, the name is marketing.

6. Is it counted once?

A single search can see the same business twice, a fraction of a second apart, when two points on the map overlap. Compare both sightings with the same earlier check and one review surge turns into two rows. Ours can't do that: an event is unique to the pair of observations it came from. Duplicates matter more than they look, because they inflate every count you'd use to decide whether a segment is worth working.

A seventh, for removals

Things appearing on a website is reliable. Things disappearing isn't. Cookie walls, A/B tests and half-rendered pages all make tools look absent for one read. A removal should need two consecutive misses before it counts, and ours does. If the tool comes back on the next check, the removal never happened.

Questions to put to any signal vendor

• What happens when your detection improves between two checks of the same business?
• What do you do with a check that failed? Can it be the "before" of a change?
• How many misses does a removal need?
• Is a numeric change sized against the business, and against the time between checks?
• When you compare two values, do you compare the thing or its label?
• For each signal name, what exactly was observed?

We hold our own signals to the same list. The thinking behind it is the same idea as keeping "no data" and "no" apart, stretched across time.