Open an AI Max campaign, sum the cost of every row in its search terms report, and compare that to the campaign’s total cost. They will not match — often not even close. Then look at the conversions the campaign claims and ask how many would have happened anyway on the exact-match keyword sitting right next to it. Three numbers that used to agree — itemised query cost, campaign total, and attributed conversions — now pull in different directions, and the account manager staring at the mismatch has to decide which one to act on before adding a single negative.
This is not a reporting bug you can file and forget. Google has been clarifying AI Max attribution discrepancies as advertisers discover search-term reporting anomalies, and the clarification amounts to: the gaps are how the system works. This post is about reconciling the three numbers you are actually shown — what is double-counted, what is hidden, and which figure to trust when you sit down to negate. It builds on how AI-interpreted terms change negation and the measurement discipline in the search terms visibility ratio.
The three numbers that no longer agree
Under classic exact and phrase match, the search terms report was close to a ledger: the queries you could see accounted for most of what you spent, and a conversion attributed to a keyword was a conversion that keyword’s query earned. AI Max breaks that tidiness in three places at once. The itemised query cost under-sums because expansion spawns rare one-off queries that fall below the reporting threshold. The campaign total is the only complete figure, but it tells you nothing about which queries drove it. And the conversion column over-claims, because AI Max expands into territory your existing keywords already covered and takes credit on the way through.
Treating these as errors to reconcile to zero is the trap. They are not inconsistent data entry; they are three different views of a system that no longer maps spend to strings. The job is not to force them to agree but to know what each one honestly tells you: the campaign total is the real money, the visible query rows are the on-theme-or-not signal, and the conversion attribution is the number most likely to be borrowed from somewhere else. Once you stop expecting them to reconcile, each becomes useful for exactly one decision instead of collectively confusing.
Where the spend hides: the itemised total under-sums
The first gap is between the summed search-terms cost and the campaign total, and it opens because AI Max is an expansion engine. One keyword becomes a long tail of queries Google predicts will convert, and most of that tail is individually too rare to clear the privacy volume threshold, so it lands in the anonymous “Other search terms” bucket. The more aggressively AI Max expands, the larger that bucket grows, and the smaller the share of spend you can itemise. This is the same visibility problem broad match created, now amplified: the queries doing the spending are precisely the ones too infrequent to show up where you could negate them one by one.
The practical move is to measure the gap rather than resent it. Sum the cost column across every visible row, divide by the campaign total, and you have the share of AI Max spend you can actually see and act on. When that share is high, the report is still a usable to-do list; when it is low, most of your money is escaping into queries you will never read, and single-query negatives are aimed at the wrong third of the account. On a low-visibility AI Max campaign the leverage shifts to pattern and n-gram negatives that block a recurring fragment across hundreds of invisible one-offs, because a negative you write against a fragment reaches queries a negative against a single string never will.
Where the conversions inflate: borrowed credit
The second gap is the one that flatters AI Max: its conversion column takes credit for outcomes your existing keywords would have won. Brad Geddes of Adalysis documented the mechanism plainly — AI Max “claims credit for conversions that would have occurred through existing exact and phrase match keywords” — because it treats every keyword as broad regardless of the match type you set, and expands into queries your tight keywords already served. So when a query converts, AI Max can report it even though the exact-match term two rows down would have captured the same user for less. The column is not lying about the conversion; it is lying about the counterfactual.
This is why the conversion figure is the worst basis for a negation or budget decision, and why so many “AI Max is winning” reads collapse under scrutiny. Independent testing across more than 250 campaigns found AI Max delivering lower ROAS than traditional match types, with less than half of search terms showing keyword-level matching. If you judge AI Max on its own conversion column you conclude it is carrying the account; if you judge it on incremental lift against the keywords it displaced, the picture is far more modest. The honest comparison is never AI Max versus zero — it is AI Max versus what the exact and phrase terms it overran were already delivering.
Which number to trust before you negate
Negate on cost and query relevance, not on attributed conversions. Cost against a visible row is money that irreversibly left the account and is the same whether or not AI Max interpreted the query; it is the one figure with no incentive to flatter itself. The query string — even when it is an interpreted-intent label rather than the literal search — tells you whether the traffic belongs to your business. Sort the search terms report by cost descending, read each query, and negate the expensive off-theme ones without regard to the conversions the row claims. A row that spent real money on an unrelated query is waste even if AI Max tagged a conversion to it, because that conversion may be borrowed from an adjacent term.
The conversion column still has a use, but a narrow one: it is a flag, not a verdict. A high-cost row with a suspiciously clean conversion is a prompt to check whether an exact-match keyword nearby covers the same intent — if it does, the conversion is probably double-counted and you can negate the AI-Max expansion to force the traffic back onto the term you control. This is the same discipline as deciding how many clicks to wait before negating: act on the reliable signal (cost accumulating on an off-theme query) and treat the unreliable one (attributed conversions) as a hint that needs corroboration, never as the trigger by itself.
Separating what AI Max drove from what it claimed
To judge AI Max fairly you have to compare periods, not columns. The campaign’s own conversion number in isolation cannot separate incremental traffic from claimed traffic, because the double-count is baked into it. What can separate them is a before-and-after at the account level: hold everything else steady, and watch what total conversions and total cost do when AI Max expansion is switched on versus a comparable baseline. If account conversions barely move while AI Max’s own column swells, you are watching credit migrate from your existing keywords into the AI Max line, not new demand being captured. That account-level read is the truthful one, and it frequently disagrees with the campaign-level story.
Segmentation makes the comparison sharper. Apply match-type and network segments to the search terms report so you can see how much of the traffic is even keyword-attributable, and lean on the segment-and-negate workflow to isolate the expansion-driven queries from the ones your keywords already owned. For accounts that also run Performance Max, the same reconciliation problem shows up across channels, which is why the Performance Max search terms report deserves the identical treatment: measure visible spend, negate on cost and relevance, and never let a channel’s self-reported conversions stand in for its incremental contribution.
Why this is permanent, and the workflow that survives it
None of this is a defect awaiting a patch. Google has been explicit that AI Max search-term matching relies on inferred intent rather than raw text, because AI Mode, AI Overviews and Google Lens each process input in ways that do not map cleanly onto the string-based model Google Ads was built on. The hidden-spend gap and the interpreted-intent labels are two symptoms of the same structural shift, and the shift is the product direction, not a bug. Waiting for the numbers to start reconciling again is waiting for a world Google is deliberately leaving behind. The reporting you have is the reporting you will have.
So the workflow has to assume the gaps. Measure the visible share of spend and re-measure it after every structural change, including the September 2026 auto-upgrade covered in the AI Max scope check. Negate on cost and query relevance, using pattern negatives where visibility is low. Judge AI Max on incremental, account-level results rather than its own conversion column. Do that, and the three numbers that no longer agree stop being a source of confusion and become three separate, honest inputs — one for how much you spent, one for whether it was on-theme, and one weak signal you corroborate before you trust. For the full reading routine, the pillar workflow in reading the Google Ads search terms report ties these habits together.