What to measure in the first 90 days of any marketing engagement

The first ninety days is where both sides set expectations they will be held to. Get it wrong and you spend month four arguing about whether anything happened.

The specific failure is measuring the wrong thing too early. An owner watching revenue weekly in month one will conclude the engagement failed before the first piece of work has had time to be indexed, let alone found.

What can honestly be measured in ninety days?

Three layers, and they arrive in this order.

Layer one, weeks one to two: did the plumbing get built? Conversion tracking live and verified with a test submission. Call tracking live. Analytics access in your own account, not the agency's. Search Console verified under your ownership. A baseline captured and dated. This layer is binary. It is done or it is not, and it is the fairest thing to hold a vendor to in month one.

Layer two, weeks three to eight: did the work ship? Pages published, listings corrected, profiles claimed, campaigns launched, technical fixes deployed. Countable, checkable, and independent of outcome. A vendor cannot control whether Google responds. They can control whether the work exists.

Layer three, weeks six to twelve: are leading indicators moving? Impressions in Search Console. Queries you now appear for that you did not before. Calls. Inquiries. Profile views and direction requests. These move earlier than revenue and they move in the right direction first if the work is working.

Revenue is a layer four question and ninety days is usually too early to answer it. Say that at the start, in writing, and the relationship gets easier.

Which layer three numbers are readable, and which are not?

Base size decides it, and the arithmetic is worth doing before the review rather than during it.

Impressions. Baseline of 12,400 non-brand impressions a month, captured before the work started. Month three comes in at 18,600. That is a 50% increase, built on a base in the thousands, and the absolute gain is 6,200 impressions. A move that size on that base is outside any plausible month to month spread. It is readable.

Inquiries. Baseline of 8 a month, ranging from 5 to 12 across the prior twelve months. Month three comes in at 11. Written as a percentage that is a 37.5%derived improvement, which is the number that will appear on the report. Written honestly it is three inquiries, and 11 sits inside a range you already knew about. It is not readable.

Conversion rate. Do not attempt it, and do the arithmetic once so you can say why. Sample size is not a traffic threshold. It is a function of how noisy the metric is and how small a change you want to catch, which is exactly how Kohavi, Deng, Longbotham and Xu state it in their KDD 2014 paper: "Formulas for minimum sample size given the metric's variance and sensitivity (the amount of change one wants to detect) provide one lower bound."

For a conversion rate the standard two proportion version fits on one line. Sessions per variant equals 16 times p times (1 minus p), divided by the square of the absolute improvement you want to detect, where p is your current rate. The 16 is 2(1.96 + 0.84)² rounded, with 1.96 for 95% confidence and 0.84 for 80% power. Put 600 monthly visits and a 3% conversion rate through it. Catching a move from 3% to 3.6% needs about 12,900 sessions per variant (derived), and at 300 sessions per variant a month that is roughly 43 months (derived). Season, spend and offer will all have changed several times over.

If somebody has told you the number is 10,000 monthly visitors, that figure has a traceable origin and it is not a study. It appears as reference [29] in the same KDD 2014 paper, where it resolves to a QuickSprout blog post by Neil Patel dated 14 January 2013, and the four authors qualify it in the sentence they cite it in: "the guidance should be refined to the metrics of interest."

The rule that comes out of it. In ninety days, report impressions and queries as evidence, report inquiries and calls as counts rather than percentages, and report conversion rate as not yet measurable. A percentage calculated on a base you could count on your hands is not a finding, it is a presentation choice.

What should be on the day one checklist?

The items that become impossible to do later.

Ownership of every account, in your name, with the vendor added as a user. Not the reverse. This is the single most consequential item on the list and the one most often skipped in the rush to start, which is why the question of what data you own and can take with you deserves settling before the first invoice.

A dated baseline export of everything you will later be judged on.

A written statement of what success looks like at ninety days, at six months, and at twelve, agreed by both sides. Specific numbers where the numbers are measurable, and an honest note where they are not.

An agreed source per metric, so month four is not spent arguing about which dashboard is real.

Server access or log access if anyone will be doing technical work. Logs are the one measurement in this whole category that is fully deterministic, and reading them for crawler activity tells you whether search and AI systems are actually fetching what was published. It also catches the most common invisible failure in this category, which is a CDN or security plugin quietly returning a 403 to legitimate crawlers. That failure is invisible in robots.txt and obvious in logs, and nothing else on the day one list is as unambiguous.

How do I read the statistics in the proposal?

Trace one of them before you sign, and pick the biggest one. Here is the trace for the number most likely to be in an AI search proposal this year.

The claim is that generative engine optimization produces around 40% more AI visibility, and that it is peer reviewed. The paper is real. Aggarwal and colleagues published "GEO: Generative Engine Optimization," accepted to KDD 2024, and it is a genuine piece of research.

What the experiment actually did is the part that never travels with the number. The authors fetched the top five Google results for a query, fed the cleaned text of those five sources to GPT-3.5-turbo with a fixed prompt, and had it synthesise a cited answer. The benchmark was 10,000 queries, 80% of them informational. The metric was Position-Adjusted Word Count. Against a no-optimization baseline of 19.3, Quotation Addition scored 27.2, and 27.2 over 19.3 is where the 41% comes from.

So the honest sentence is: one tactic produced a 41% lift on a word-count metric, inside a simulation of a search engine, over five fixed sources, on non-local informational queries, in 2023 and 2024. It is not traffic. It is not citations in ChatGPT or AI Overviews. And nothing in the benchmark is a local commercial query, so the paper has not been validated for local intent at all.

The findings that survive are useful and less exciting. Adding verifiable statistics, direct quotations and inline citations made the model more likely to lean on a page. Keyword stuffing actively hurt, scoring 17.7 against the 19.3 baseline. An authoritative tone produced no significant improvement, and the authors say so directly.

Run that trace on whichever number is doing the most work in your proposal. It takes twenty minutes and it tells you more about the vendor than the reference check will.

Which early numbers mislead?

Several, and they mislead in predictable directions.

Rankings in the first month. Positions move constantly, vary by location and device, and mean little without volume attached. A screenshot of one improved ranking is the oldest report filler there is.

Traffic spikes. A new page, a social post, or a bot wave produces a bump that looks like progress. Check the source before you celebrate.

Small conversion changes. Covered above. Going from eight inquiries to eleven is three inquiries.

Any single AI visibility score. These systems are non-deterministic. Researchers at the University of St. Gallen ran daily prompts across four AI engines over a 45 day window plus repeated same day runs, and found that identical prompts submitted on the same day shared only 32% to 43% of their cited sources. Their prescription is at least seven runs per prompt per day for a brand visibility figure, and reporting on a two to four week rolling aggregate rather than week over week. A number without a sampling method behind it is an opinion with a decimal point.

The honest framing to insist on: know which numbers are counted, which are estimated, and which nobody can measure. A report that mixes all three without labelling them is doing the mixing on purpose.

How do I run the ninety day review?

Like an onboarding review, not a performance trial.

Bring the baseline document. Bring the written success definition. Go layer by layer. Was the plumbing built. Did the work ship. Are leading indicators moving. Only then, what does revenue look like, with the caveat already agreed.

Ask one question that reveals more than any metric: what did you learn about my business that you did not know at the start. A vendor doing real work will have found something specific, an unexpected query, a broken page, a competitor pattern. A vendor filling hours will answer in generalities.

Then read the report itself with attention. Reading a monthly marketing report critically is a skill worth developing before the ninety day mark, not after.

The structure is the same one that works for new staff. The first two weeks of a new kitchen hire decide most of what follows, because that is when standards get set or quietly abandoned. An agency engagement behaves identically.

What to do this week

If your engagement has not started, write the ninety day success definition yourself before the kickoff call. Two paragraphs. Bring it to the meeting.

If you are already thirty days in and none of the plumbing exists, that is your finding. Fix it this week rather than waiting for the review.

Verify one conversion yourself. Submit your own form, tap your own phone number, and confirm it appears in the reporting. This takes four minutes and it catches the most expensive failure in the category.

Pick the largest statistic in the proposal and spend twenty minutes finding its primary source. Not a blog citing it, the document itself.

And if the early data shows traffic arriving but nothing converting, deal with that before adding more channels. Fixing the first conversion problem is worth more than being slightly better everywhere, especially in a first quarter when budget is tight.

Questions about your first ninety days?

Email me at eric@seod.com with the deliverables list from a proposal you are considering or an engagement you have just started. Paste it in. I will mark which items can honestly be measured inside ninety days, which will not show until month six, and which are not measurable at all.

I answer these myself and I will not use it as an excuse to pitch against whoever wrote the proposal. If the plan is a good one, that is what you will hear.

More on measurement sits in the analytics library.