A fair ninety day test of a marketing agency starts with the baseline you take before day one
Three things get fixed in week zero, or day ninety becomes an argument about which numbers to look at.
Published by AI Knows Us · Updated 22 October 2026
Ninety days is long enough to judge a marketing agency fairly, and only if three things are fixed before the first invoice. An exported baseline of every measure you intend to use. Two or three measures, chosen in advance and written down. A decision rule that says what continue, change and stop each look like. Fix those in week zero and day ninety is a short meeting. Skip them and day ninety becomes a negotiation about which chart to open, which the party with more charts wins.
The week zero baseline pack
Take these exports and save them with the date in the filename, before the agency touches anything. They take an afternoon and they are the whole foundation of the test.
- Search Console, the full available history. Clicks, impressions, average position and queries. Export it rather than relying on the interface, because the window inside the tool rolls forward and your comparison point moves with it.
- Analytics, twelve months by channel. Sessions, conversions and the channel split, so a later claim about organic growth has something behind it.
- Enquiry count by month, twelve months, counted by you. From your inbox, your call log and your CRM, not from a platform.
- Ten commercial phrases and where you rank today. Ten, chosen because somebody buying would type them, recorded with the date and the device.
- The site as it stands. A crawl, a list of indexed URLs, and the page count. This is what makes a change log checkable later.
- Advertising spend and account change history. So an improvement in cost per acquisition can be separated from a change in budget.
Send the pack to the agency as well. A supplier that is being measured on a baseline it has seen behaves differently from one being measured on a baseline it discovers at day ninety.
What genuinely moves inside ninety days, and what does not
| Work | What ninety days can show |
|---|---|
| Technical fixes, indexing, site speed, broken templates | Full effect. These are the fastest wins and they should appear in weeks two to six |
| Paid search and paid social | Full effect on cost per enquiry, given enough volume. Judge the cost, not the spend |
| Conversion work on existing pages | Measurable, if the page already has traffic. A test needs visitors to run on |
| Local presence and the Business Profile | Visible movement in calls and direction requests |
| New pages and new content | Indexing and impressions, usually before clicks. A page can be collecting impressions at a poor position and still be working |
| Domain authority and links | Little. This is the one that compounds over quarters, so judge the activity and the quality rather than the outcome |
What a working quarter actually looks like on a site with history
Our founder's legal search company, clawlaw.in, is published as a worked case on our proof page, from a Google Search Console export taken on 14 October 2026. Over the window from 13 July to 12 October 2026, 92 days with data, the site recorded 18,537 clicks from 929,405 impressions. Clicks a day moved from 59 on the first day to 349 on the last. Month by month the daily averages ran 59 clicks and 2,253 impressions in July, 154 clicks and 7,276 impressions in August, 282 clicks and 14,703 impressions in September, and 349 clicks and 18,331 impressions across the first 12 days of October.
That domain had history and existing pages before the window opened, so those are growth figures on a working site rather than a launch. The shape is the useful part. Impressions moved first and by a wide margin, clicks followed, and the curve was still climbing at the end of the window. A quarter that looks flat in month one and steep in month three is the normal shape of this work, which is exactly why the day thirty conversation should not be a verdict.
The three checkpoints, and what each one is for
Day thirty, an access and progress check
Not a results meeting. Three questions. Does the agency have everything it asked for, or is it waiting on you. Has anything shipped, with dates and links. Has the baseline been agreed in writing by both sides.
Most disappointing engagements are visibly off track here, and the symptom is almost never a bad number. It is that nothing has shipped and the reason is an approval sitting with somebody on your side.
Day sixty, the first honest read
Technical and paid work should be showing. New pages should be indexed, which you check yourself rather than taking on trust. Impressions on non branded queries should have moved even if clicks have not.
This is the right moment to change scope. Two months in, both sides know something they did not know at the start, and a scope that has not changed at all by day sixty usually means nobody has learned anything.
Day ninety, the decision
Compare against the week zero exports, on the measures agreed in week zero, with branded search filtered out. Then apply the rule you wrote down.
The decision rule, written before you need it
- Continue when the agreed measures have moved in the right direction, the change log is full and dated, and the next quarter's plan names dependencies on both sides.
- Change the scope when the work shipped and the measures did not move. That is usually a targeting problem rather than an effort problem, and it is the most common honest outcome of a first quarter.
- Stop when the change log is thin and the reason is on the agency's side, or when you have asked twice for a figure traced back to one of your own systems and not received it.
Write those three paragraphs into the contract alongside the notice period. A decision rule agreed in advance removes the one thing that makes day ninety unpleasant, which is both parties arguing for the test they would pass.
What not to judge at ninety days
- Brand awareness. No measure available to a mid sized business reads it in a quarter.
- Link building outcomes. Judge the links themselves and the sites they sit on.
- A new service line with no existing demand. You are creating a market, and the clock for that is different.
- Anything affected by a seasonal peak inside the window. Note the peak, compare to the same peak last year, and say so in the report.
Questions about judging an agency at ninety days
Is ninety days long enough to judge SEO?
It is long enough to judge the work and the direction, and not the full outcome. Technical fixes and indexing land inside the window, impressions move, and clicks usually follow in the quarter after. Judge delivery and leading measures at ninety days and revenue at six to nine months.
What if the agency wants a twelve month minimum term?
Agree the term and put a ninety day review with named exit criteria inside it. A supplier confident in its first quarter will accept a review it can pass.
We never took a baseline and we are at day eighty. What now?
Search Console holds a long history, so export it today and use the three months before the engagement as your baseline. Your own enquiry records go back further than any platform. The test is still worth running.
Should the agency choose the measures?
It should propose them and you should approve them. A proposal that avoids enquiries and revenue in favour of impressions and rankings is useful information about what the agency expects to achieve.
How many measures should we use?
Two or three. More than that and every quarter produces a mixed picture, which always resolves in favour of continuing.
The measure almost nobody baselines in week zero
The pack above captures everything Google and the ad platforms will tell you later. It does not capture what an assistant said about your business on the day the engagement started, and that is now part of how buyers build a shortlist. Taking that reading in week zero costs an afternoon and makes the day ninety comparison possible at all, because a count with no before cannot show movement. AI Knows Us takes that baseline across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews and repeats it, and what a British before and after can honestly read from a short window is set out in our note on a British before and after test.