Diagnose before you do anything
Most wasted effort in this field comes from treating AI visibility as one problem. It is at least three, they look similar from the outside, and they need different work. Before committing a budget, take a baseline measurement and decide which of them describes you.
- Unknown: you are rarely or never named in category answers. The bottleneck is usually evidence — thin content, little external corroboration, an unclear entity.
- Misdescribed: you are named but the description is wrong or incomplete. The bottleneck is contradictory or stale sources, and it is the cheapest of the three to fix.
- Mentioned but not recommended: you appear in the list without being put forward. The bottleneck is differentiation — nothing in the available evidence says what you are specifically best for.
Establish the baseline properly
A baseline is not a screenshot of one helpful answer. It is a fixed prompt set, sampled repeatedly across the platforms your buyers use, recorded for presence, prominence, recommendation, accuracy, citations and who else was named. Our article on measuring AI visibility sets out the method; the important thing is that the same instrument is used every time, or improvement becomes unprovable.
Write the baseline down, including the number of samples and dates. In three months you will want to know whether a change is real or ordinary variation, and without a recorded starting point that argument is unwinnable — particularly if somebody is asking what the investment achieved.
Days 1–30: remove the contradictions
The first month is correction, not creation. Settle your core facts, then make every source agree: your own pages, third-party listings, professional registers, social profiles, old landing pages and anything else still indexed. Remove or update stale pages rather than leaving them to contradict the current ones.
In the same window, fix the technical basics that silently block everything else. Check that your pages render their text without a script, that structured data exists on the commercially important pages and matches the visible content, that your sitemap is current, and that your robots.txt reflects a deliberate decision about which AI crawlers may fetch your pages. Finding a retrieval crawler blocked by an inherited rule is common and is the cheapest win available.
If your diagnosis was "misdescribed", this month may be most of your programme. Re-measure at the end of it before spending anything further.
Days 31–60: make yourself quotable on the questions that matter
With the facts settled, choose the five to ten buyer questions with genuine commercial value — the ones that precede a real enquiry, not the ones with the most search volume — and write properly on each. Answer in the first two sentences, use a question-shaped heading, and make each section able to stand alone if lifted out.
Attack differentiation in the same pass, because this is the work that turns a mention into a recommendation. State what you are specifically best at, for whom, with what evidence, and state what you do not do. Vague breadth gives an assistant nothing to recommend you for; a narrow, documented specialism gives it a reason to name you when a matching question arrives.
The editorial craft here is covered in our explainer on answer engine optimisation, and the full tactical checklist in our guide to appearing in AI search. Resist publishing volume: a dozen thin pages perform worse than three unambiguous ones.
Days 61–90: corroboration and the first honest re-measurement
Independent agreement is the slowest lever and the one most often skipped. Use this window to get your entries on legitimate industry directories and registers current and identical, to contribute real expertise where your industry publishes it, and to make sure anything genuinely notable is documented somewhere you do not own. Encourage authentic reviews through the platforms you already use; never buy or fabricate them.
Then re-run the baseline — same prompts, same platforms, same number of samples — and compare distributions rather than anecdotes. Expect a messy picture: accuracy usually improves first because it depends on sources you control, inclusion moves more slowly, and recommendation is the last to shift because it depends on external agreement.
How to read the results without fooling yourself
Two disciplines keep this honest. First, compare like with like: a changed prompt set invalidates the comparison, and so does switching platforms between measurement rounds. Second, attribute cautiously. Platforms update their models and retrieval systems on their own schedule, so a jump or a drop may have nothing to do with your work. Note model versions and dates, and say "we cannot separate these effects" when that is the truth.
Also keep expectations calibrated on timing. Retrieval-driven answers can reflect changes within weeks. Claims held in a model’s internal knowledge can take far longer, and nobody outside the platforms can tell you how long — a limitation explained in how ChatGPT chooses which businesses to recommend.
What not to spend money on
A short list, offered plainly. Guaranteed inclusion in AI answers: nobody can deliver it. Bulk directory submissions or paid mention networks: the exact pattern platforms have spent years learning to discount. Tools that report a confident AI visibility percentage without showing you the prompt set, platforms, dates and sample size: the number is unverifiable. Content written at volume to cover every phrasing of a question: clarity beats coverage when the system is summarising meaning.
What does repay investment is unglamorous: accurate facts, consistently stated, corroborated by sources you do not own, on a site that retrieval systems can read — measured with the same instrument each time. That is the whole programme, and it is the version of it we would defend to a sceptical finance director.
Last reviewed 28 August 2026. AI Says is not affiliated with or endorsed by OpenAI, Google, Anthropic, Perplexity or other AI and search providers; platform names are used descriptively.