Nine questions that expose whether an agency is current

You cannot assess a marketing partner by looking at their work. Everyone's case studies look good, everyone's deck is beautiful, and the results were produced by an environment that has since changed.

You can assess them with nine questions. Each one has a right answer, each answer is verifiable, and none of them require you to be a specialist. Use them on an incumbent as easily as on a pitch.

1. What happens to our account in September 2026?

Right answer: Google begins auto-upgrading campaigns that use Automatically Created Assets or the campaign-level broad match setting to AI Max, and stops new Dynamic Search Ads campaigns. February 2027 is the mandatory sweep for the rest. They should know which of your campaigns are affected without checking.

Wrong answer: anything that only mentions DSA. Plenty of teams have the February date in a calendar and have missed that the assets and broad match trigger lands first.

2. How many attribution models are available to us?

Right answer: two. Last click and data-driven. The others were removed from new conversion actions in June 2023 and force-migrated in September 2023.

If someone offers to build you a custom position-based model in the platform, they have not opened it recently.

3. Are we eligible for data-driven attribution in GA4, and how do you know?

Right answer: they cite the thresholds. At least 400 conversions for the specific conversion action and at least 20,000 total conversions across all actions inside the lookback window. Then they tell you whether you pass.

Most mid-sized businesses with a considered sale do not pass on their primary action, which means the model they are quoting to you is last click wearing a different name.

4. Design me a holdout for our largest channel.

Right answer: they name the unit, regions or postcode groups rather than two cities, the control share of roughly 20 to 30 per cent of baseline conversions, the matching method of pre-period trend rather than population, the minimum duration of at least one full purchase cycle, and the two things that ruin it, spillover from broad media and branded search leaking the effect into another channel.

Wrong answer: an offer to prove incrementality using platform-reported lift only.

5. What did the 15 June 2026 consent change do to our data?

Right answer: ad_storage became the sole control over whether GA4 shares cookies and identifiers with linked Google Ads accounts, and the Google Signals toggle now only governs signed-in behavioural reporting. Accounts can now over-collect or silently lose Ads data depending on how they were set. The audit is done in DevTools by comparing the gcs and gcd parameters on a reject versus an accept, not in Tag Diagnostics, which lags two to three days.

This is the single best question in the list. It is recent, it is technical, and it cannot be bluffed.

6. Should we publish an llms.txt file?

Right answer: no, Google Search does not use them. Google said so in its May 2026 guide on optimising for generative AI features. The same guide says structured data is not required for generative AI search, and there is no requirement to break content into tiny chunks.

Anyone selling you an AI-search package built on those three items is selling a 2024 hypothesis.

7. What is our engaged-view and view-through window set to, and why?

Right answer: a specific number, tied to your sales cycle, with a note of when it was last changed. Generous view-through windows make video and display look extraordinary, and a window change makes two periods incomparable.

If nobody knows, your video reporting is decoration.

8. How much creative do you produce a month, and how do you retire it?

Right answer: a number of genuinely distinct concepts, not variations, plus a fixed retirement schedule. Both platforms now buy audiences through creative rather than filters, so production capacity is a media variable.

Wrong answer: we test headlines.

9. What in our marketing does not work?

Right answer: something specific, costed, and inconvenient for them to say. A channel they think you should cut. A page that is losing you money. A limitation of their own approach.

A partner who cannot name a failure has either not looked or will not tell you.

How to read the answers

You are not testing recall. You are testing whether the person in front of you touches the account and reads the platform documentation, or whether they are repeating a training they did three years ago.

Two more things worth doing. Ask who will actually do the work and get the name. Then ask what happens in month four, because most partnerships are excellent for ninety days and average afterwards, and the honest ones will tell you how they handle that.

Ask them today, before the September date makes the first question retrospective.