Bidding Update Impact

Did Google's 17 August 2026 bidding change actually hit this account — and what did it cost?

On 17 August 2026 Google stopped letting budget-limited campaigns on a target-based bid strategy overperform their target. A tCPA of €10 that had been quietly delivering €5 now drifts toward €10. Spend is capped either way; what changes is what that spend buys.

Half the industry is currently answering this question with a before/after spreadsheet. That answer is usually wrong.

Why the obvious approach fails

From a real account (PMax, tROAS 6.0, budget-constrained):

| Window | Spend | Conv. value | ROAS | |---|---|---|---| | 10–16 Aug | 34,577 | 155,915 | 4.51 | | 17–23 Aug | 34,267 | 128,933 | 3.76 |

Reads like a 17% collapse. It isn't. That account's own conversion-lag curve shows only 79.8% of conversion value has matured by day 7 — the post window was still filling in. Corrected, it lands at ~4.7: up, not down.

The naive comparison inverted the conclusion.

On synthetic data with a known answer, the gap is starker still: in a scenario with zero injected effect but a normal seasonal dip, raw ROAS reads 4.99 → 3.74, a 25% apparent collapse. This skill returns NOT_YET_DETECTED and a difference-in-differences of −0.0%.

What it does

1. Classifies every campaign into treatment and three control groups, handling the four different fields a bid target can hide in and the campaign types Google excluded. (A live account had an App campaign on Target CPA and flagged budget-constrained — a textbook false positive. App is out of scope.)
2. Corrects for conversion lag using the account's own curve, so a fresh window isn't compared against a matured one.
3. Difference-in-differences against campaigns the update provably cannot have touched, so seasonality cancels instead of being reported as impact.
4. Excludes contaminated campaigns — anything a human re-budgeted or retargeted mid-window gets no verdict, and says who changed it.
5. Recommends targets for campaigns still at risk. It never writes to the account.

Output: a self-contained report.html, plus analysis.json for auditing and raw_pull.json as a permanent snapshot.

Verdicts

| | | |---|---| | AFFECTED | Treatment converged toward target; control didn't. Real signal. | | NOT_YET_DETECTED | No divergence found. The rollout is gradual — not yet, or not at all. Never "unaffected". | | TOO_EARLY | Post window too short or too immature. No verdict is honest here. | | CONFOUNDED | Too many mid-window edits to attribute anything. |

Even on a null verdict the report gives the exposure figure — spend sitting in campaigns that were overperforming and are therefore still at risk whenever the rollout reaches them. That number is actionable regardless.

Run it

# demo, no account needed
python scripts/sample_data.py --scenario affected --out raw_pull.json
python scripts/analyze.py --raw raw_pull.json --out analysis.json
python scripts/build_report.py --data analysis.json --out report.html

Against a real account, replace step 1 with the five GAQL pulls in reference/gaql-queries.md (all validated live) and follow SKILL.md.

Scenarios: affected, null, too-early, confounded — each with known ground truth, which is how the engine is tested.

⚠️ Time limit

change_event reaches back 30 days only. From roughly 17 September 2026, change history around the switch date is gone permanently — after that there is no way to prove whether a target was human-edited or moved by the rollout. Capture raw_pull.json for any account you might want to analyse before then.

Files

SKILL.md                      workflow the agent follows
reference/the-change.md       what Google actually announced, with sources
reference/methodology.md      why the naive test fails, and the honesty rules
reference/gaql-queries.md     validated GAQL + the field traps
reference/data-schema.md      raw_pull.json contract
scripts/analyze.py            classification, lag correction, DiD, verdict
scripts/build_report.py       HTML renderer
scripts/sample_data.py        ground-truth fixtures for tests and demos
assets/                       example reports

Requires Python 3.9+ and the googleAdsServer MCP. No third-party packages. Read-only throughout.