Negative keywords now cost you conversions

Google now tells advertisers to use negative keywords sparingly. That instruction sits inside its AI Max guidance, and it inverts the single most common optimisation ritual in paid search: open the search terms report, sort by cost, negate anything that looks wrong.

The instruction is not laziness on Google's part. It follows from how matching works now, and once you see the mechanism the advice stops feeling like a surrender of control.

What query fan-out did to the search terms report

Search used to be close to literal. A person typed a string, your keyword matched it more or less loosely, and the search terms report showed you the string. Negating a bad string removed a bad auction entry. Clean, one to one.

That relationship is gone. Search term matching in AI Max, and increasingly the broader Search system, expands a query into related concepts and matches against page content and intent rather than a token list. The report now shows you a record of machine-inferred intent, not a list of things people typed that you can safely delete.

So when you negate a term, you are rarely removing one bad query. You are removing a signal from a cluster the match layer was using to reach a whole neighbourhood of queries, most of which you never saw individually because they each carried three impressions.

That is the tax. It is invisible, because the conversions you did not get do not appear anywhere.

What still deserves a negative

Restraint is not abolition. Some things should never be bought, and no amount of modelling changes that.

Negate these without hesitation:

• Job seeker intent. Careers, jobs, salary, hiring, apprenticeship. This is the single largest waste category for medium and large employers, because your brand name is genuinely popular with job seekers.
• Existing customer support intent. Login, portal, invoice, warranty claim, contact number. You are paying to serve people who already pay you.
• Wrong product line. If you sell commercial and not residential, negate the residential qualifiers explicitly. The model cannot infer a commercial-only mandate from a website that talks about both.
• Regulated or reputational terms. Anything your legal or compliance team would object to appearing next to your brand.
• Other companies' brand names, where you have decided not to bid on them.

Those five categories are structural. They describe a permanent boundary of your business, not a query that underperformed last fortnight.

What no longer deserves a negative

This is the harder half.

Stop negating on the basis of a single expensive click with no conversion. At the query level, almost nothing has enough data to be judged. A term with 4 clicks and 0 conversions is statistically identical to a term with 4 clicks and 1 conversion. You are reading noise and turning it into a permanent rule.

Stop negating terms that look odd but sit inside a converting cluster. If a group of long-tail queries around a topic is converting at target and one member of that group looks strange to you, the strange one is often doing work you cannot see, because it is teaching the match layer where the boundary of your relevance sits.

Stop using single-word broad negatives. This is the most damaging habit in the list. A phrase-match negative on the word free blocks freehold, freestanding and free-standing. A negative on the word cheap blocks nothing useful and often collides with legitimate value-led queries in trades and manufacturing. One-word negatives feel efficient and cause the most collateral damage per character typed.

A process that survives the change

Replace the weekly negation sweep with something slower and better documented.

1. 1.Review fortnightly, not weekly. This matches Google's own recommended cadence for AI Max search terms and item groups, and it gives every query cluster a chance to accumulate enough data to say something.
2. 2.Group before you judge. Export search terms and cluster them by theme, not by cost. Judge the theme. A theme with 300 clicks and 2 conversions is a real decision. A term with 5 clicks is not.
3. 3.Quarantine before you commit. Keep a shared list called Quarantine. New negatives go there with a date and a one-line reason. Review it monthly. Terms that were added on thin evidence and never revisited are the reason accounts slowly suffocate.
4. 4.Prefer exact-match negatives. Phrase and broad negatives are where the accidental blocking happens. Exact costs you a few more rows and removes the surprises.
5. 5.Keep permanent structural negatives in an account-level shared list. Keep experimental ones at campaign level. Then when someone asks why volume dropped, the two are separable.

The measurement problem, named honestly

You cannot easily prove the cost of over-negation, because the counterfactual is missing. The only clean read is to build two near-identical campaigns, apply your standard negative list to one and only the structural categories to the other, and run them long enough to compare volume and blended efficiency. That takes budget and about six weeks. Most teams will not do it.

The next best thing is to check the direction of travel. Pull your negative list count by campaign and plot it against impression share lost to rank over the last twelve months. If the list has grown steadily and your eligible impressions have shrunk, you have a candidate explanation that is worth testing.

The uncomfortable part

Negation is satisfying. It is the one lever in a heavily automated account that feels like craft. Deleting a wasteful query produces an immediate, visible saving, and nobody ever gets criticised for it.

The saving is real and the cost is deferred, which is the definition of a bad trade in a system that learns.

Audit your negative lists for one-word entries this week. That single pass usually finds the biggest unexploded charge in the account.