AI Hallucination in PPC: When the Numbers Are Made Up

AI hallucination is a language model asserting something plausible that is not true. In PPC work it has a specific, recognisable shape: a metric that was never fetched, a setting described from memory of how accounts usually look, a tidy month-on-month comparison for a month the model never saw. The numbers arrive fluent, formatted and wrong — and the tell is always the same: no tool call behind them.

The word suggests malfunction, but the mechanism is ordinary. A language model is a prediction engine, not a database; asked for your CTR without access to your account, the only thing it can produce is a statistically plausible CTR. Confabulation is what filling a gap looks like when the generator is fluent.

How hallucination happens — and what grounding changes

Ungrounded, the failure modes in ads work are predictable. Invented metrics: ask a disconnected chatbot about your wasted spend and it will discuss plausible wasted spend, sometimes with invented specifics. Misremembered configuration: descriptions of your settings drawn from the average of all accounts rather than yours. Plausible-but-wrong technicalities: GAQL queries with field names that nearly exist. Each error is fluent enough to survive a skim, which is what makes the class dangerous in a spend context.

Grounding changes the equation structurally. An agent connected to the account through tool calls does not need to imagine your data — it fetches, and the fetch is visible: tool, account, query, result, all shown in the client. For anything actually retrieved, the invented-number problem disappears, because the number has a receipt.

What grounding does not do is finish the job. A connected model can still misread what a tool returned, carry a stale figure forward through a long conversation, or quietly bridge an ungrounded gap with pattern memory. The failure shrinks from "made up" to "mishandled" — a smaller problem, and still a problem.

The receipts habit that closes the gap

The working discipline is receipts: every number that matters should trace to a visible tool call. Clients display each call an agent makes, so the check costs seconds — the metric either has a fetch behind it, with the date range you intended, or it is conversation. This is the same verification habit that applies to changes — trust the log, not the recollection, as argued in verify what your AI changed — extended to reads.

Prompt patterns push in the same direction, and they are simple enough to paste at the top of any working session:

Only report numbers returned by tool calls in this conversation.
State the date range with every metric.
If data is unavailable, say so — do not estimate.
Show which query produced each table.

Instructions like these convert the model's default helpfulness — filling gaps smoothly — into the behaviour a spend system needs: flagging gaps loudly. A fuller set of working rules is in the prompt guidelines for Google Ads AI.

The honest summary: hallucination is why an unconnected chatbot cannot manage an ad account, and grounding-plus-receipts is why a connected agent can. The difference is not a smarter model. It is an architecture in which asserting requires fetching — and a user who glances at the receipts.