The first ninety days of AI visibility work

Key takeaways

• Spend the first month on questions, a baseline and a diagnosis. Work that starts before the baseline cannot be measured against it.
• Fix owned evidence first. It is the fastest to change and everything else, including earned evidence, depends on it.
• Measure on the same questions at the end of the quarter and report what did not move as plainly as what did.

AI visibility work fails in two predictable ways. Some teams spend a quarter reading answers and producing reports, and never change anything. Others start publishing in week one, before they know where they stand, and arrive at the end of the quarter unable to say whether any of it helped.

Ninety days is long enough to avoid both. It gives you time to set a baseline you can trust, to fix the evidence that matters most and to measure the result on the same questions. It is not long enough to change everything, and this plan does not pretend otherwise. It is a plan for a first cycle that leaves you with evidence, a working method and a clear view of what to do next.

The plan assumes a small team: a marketing lead who owns it, a content owner, some time from someone who can edit the website and an hour a fortnight from sales.

Weeks	Phase	What you finish with
1 to 2	Agree the questions	A frozen, weighted buyer-question set
3 to 4	Set the baseline and diagnose	Recorded answers, three separate counts and a list of evidence gaps
5 to 8	Fix owned evidence	Published pages for the priority gaps, each logged with a date
9 to 10	Start on earned evidence	Partner, directory and review pages updated or requested
11 to 13	Measure and report	A like for like comparison, flat results included, and the next quarter's plan

Weeks one to two: agree the questions

Everything that follows depends on asking the right questions, so do not rush this.

Work with sales to collect the questions buyers actually ask, from first calls, lost deals, RFPs and support tickets. Group the wordings into prompt families, place each on the buying journey and agree a commercial weight for each before anyone reads an answer. Building a buyer-question set sets out the method in detail.

At the end of week two, freeze the set. Give it a version and a date. The comparison at the end of the quarter is only meaningful if the questions are the same ones you start with.

Agree two other things at the same time. First, which rivals you will track alongside your brand: two or three close competitors is enough. Second, who owns the plan. A programme without a named owner becomes a report.

Weeks three to four: set the baseline and diagnose

Now read the answers. Ask each question in your set and record, for every answer:

• the date and the assistant;
• whether your brand and each tracked rival were named;
• the role each played, from primary recommendation down to not present;
• the sources the answer cited, and whether any were on your domain;
• the reason the answer gave for its choice, in its own words.

Ask each question more than once. A single answer is an anecdote. Repeated answers to the same agreed questions show you which positions are stable and which come and go.

Then count. Keep the three outcomes apart: how often you were mentioned, how often your own evidence was cited and how often you received a qualified recommendation. Report each over its own base, with the counts beside the rate. Reading an AI visibility figure honestly explains why a baseline resting on a handful of questions is a direction rather than a measurement, and how to say so.

Finally, diagnose. For each answer where you were absent or named but not chosen, note what was missing: a page on the question, a specific claim, a citable source, an independent voice, a critical claim pulling the other way. Evidence gaps describes the twelve classes we use.

Weeks five to eight: fix owned evidence

This is where the work happens. Start with your own domain, for three reasons. It is the evidence you control completely. It changes fastest. And almost every other fix depends on it: a partner cannot link to a page that does not exist, and a reviewer cannot quote a capability you have never described.

Prioritise by weight and by pattern. Take the gaps that recur across your highest weighted prompt families and work through them in that order. For most teams, the first month of fixes looks something like this:

1. A page for each priority question you were absent from. Written about the question, in the buyer's words, with the specific facts an answer would need.
2. Specific claims where you were named but not chosen. If a rival won on "documented UK data residency", publish yours, with the detail a compliance lead would want.
3. Citable homes for important claims. Security, integrations, certifications and sector experience, each on a findable, dated page.
4. A technical check. Confirm that the important pages can be crawled and indexed, and that their key text is in the page rather than loaded by scripts or held in images.
5. Consistent names. One name for each product and capability, used the same way across your site, documentation and listings.

Log every change as you ship it: the date, the page, the gap it addresses and the questions it should affect. This intervention log is what lets you connect a change in the answers to a piece of work, rather than guessing at the end of the quarter.

Weeks nine to ten: start on earned evidence

Some of the most influential evidence is published by other people. You cannot write it, but you can make it accurate and likely.

• Partner and marketplace listings. Update every listing you control. An incomplete or out of date listing is easy to forget, and it can be cited like any other page.
• Directories and review sites. Make sure your profile describes the same capabilities, in the same words, as your own pages.
• Analysts and comparison authors. Where a comparison describes you inaccurately, send the owner a short, factual correction with a link to your current page.
• Critical sources. Where an outdated complaint is pulling answers away from you, publish what changed and ask the source's owner to reflect it.

Earned evidence moves slowly. Starting it in week nine means it will not all show in this quarter's measurement, and that is fine. Note what you requested and when, so next quarter's results can be read against it.

Weeks eleven to thirteen: measure and report

Wait until the owned work has been live for a few weeks, so it has had time to be found. Then ask the same frozen questions again, the same number of times, and record the answers the same way.

Compare like with like. Put the three outcomes beside their baseline values, each with its counts and its interval. Look for movement on the specific questions your work was aimed at, not only the overall figures. And check that nothing else changed underneath: if you altered the question set, the tracked rivals or the way you collect answers, say so, because the before and after are no longer directly comparable.

Then write the report a sceptical board member would trust:

• What we did. The intervention log, in short.
• What moved. The questions and outcomes that changed, with counts, and whether the change is larger than ordinary variation.
• What did not. The questions where nothing moved, stated as plainly as the ones that did.
• What we learned. Which gaps turned out to matter, and which did not.
• What next. The gaps for next quarter, in priority order, with owners.

Resist the temptation to promise that the next quarter will deliver more. Nobody outside the providers controls what an assistant recommends. What you can promise is that the evidence will keep improving and that the measurement will stay honest.

What not to do in the first ninety days

A few mistakes are common enough to name.

• Do not change the questions mid-quarter. If you must, start a new baseline.
• Do not blend the outcomes into one score. It will hide the movement that matters most.
• Do not chase every gap at once. Ten pages that close the most common gaps beat forty that close none of them well.
• Do not publish material written to manipulate a model. Thin pages stuffed with your brand name, invented reviews or hidden text are the kind of trick that depends on nobody noticing, and they put the evidence you have built at risk.
• Do not read one answer as a result. Not in week three, and not in week thirteen.

At the end of ninety days you should have a frozen question set, an honest baseline, a log of what you changed, a like for like comparison and a prioritised list of what to do next. Repeat the cycle on the same questions and the picture sharpens every quarter.