Four stages, and we will tell you if the first one is enough
The whole engagement, in the order the evidence supports. Stage one produces documents you keep whether or not we work together again.
Short answer
We establish a baseline you can argue with, fix what demonstrably blocks retrieval, work on the half of the problem you do not own, and then measure it honestly enough that a flat month reads as a flat month.
01 · Baseline
Two things happen in parallel, and both produce a document you keep whether or not you work with us again.
We sample the answers. A fixed set of local-intent prompts for your category and region, run across ChatGPT, Claude, Perplexity and Google AI Overviews. Not once: repeatedly, because a generated answer varies between identical runs and a single sample is noise. Evertune's repetition analysis puts the margin of error at 27 points after five runs of a prompt and 12 points after twelve, with returns flattening around a hundred. We tell you the sample size and the spread next to every number.
We fetch your site the way the crawlers do. JavaScript disabled, one request per page, reading only what comes back. Then robots.txt, then the CDN rules, because those are two separate places you can be blocked and most sites that think they have checked have only checked one.
- Prompt-set results with sample size and spread, and the competitive set: who gets named instead of you, and which sources those answers drew on.
- A crawler access report quoting the exact lines from robots.txt and naming any CDN rule turning a bot away.
- A server-render diff: the same pages with and without JavaScript, and a list of everything that disappears.
- The uncomfortable list: facts about your business that no source states clearly. Hours, price band, capacity, what you do not do.
At the end of this stage we tell you whether the rest is worth doing. We have told people it is not.
02 · Retrievability
The mechanical half. It is cheap, it is verifiable, and if any of it is wrong nothing else you spend money on can work, because the page is never read.
No major AI crawler executes JavaScript. Content assembled in the browser is invisible to every one of them except Google's surfaces and AppleBot. So content moves to the server, OAI-SearchBot, Claude-SearchBot, PerplexityBot and Bingbot get unblocked in both places they can be blocked, dead URLs and redirect chains get cleared, and the paragraph under each H1 gets rewritten to answer the question the page exists for.
Schema goes in as well, because it is cheap and it earns conventional rich results. Google states no special structured data is needed for its AI surfaces, and a matched study found adding JSON-LD produced no major uplift in citations anywhere. We deploy it and we describe it honestly. It is not the lever it is sold as.
03 · Off-site mentions
The slow half, and the one with the most leverage. Across roughly 75,000 brands, branded web mentions correlated with AI Overview visibility at 0.664 against 0.218 for backlinks. The authors say plainly that correlation is not causation, and so do we. It is still the strongest published signal in the field.
We do not work from a generic directory checklist. The baseline sampling already told us which domains the assistants actually cite for your category and city, and that list is the target. First your details go consistent across them, because contradictions between sources are what make a model hedge or route around you. Then the slower work of being included in the guides and roundups that came back in the sample.
This is the part we cannot put a date on. Editorial timelines belong to editors.
04 · Measurement
The same frozen prompt set, on the same cadence, logged out, on a fixed locale, with the methodology unchanged so periods are comparable. Changing the prompt set mid-engagement is the easiest way to manufacture an improvement, so when we do change it we report old and new side by side.
- Appearance rate, with the sample size and the spread. When a figure moves less than the margin, we say it did not move.
- Description accuracy. When you are named, what the assistant says about you, and whether it is right. Wrong hours in an answer cost more than absence.
- Mention versus citation, tracked separately. A source can be cited without your brand ever being named, and if you sell something the mention is the outcome that matters.
- Mechanical status. Sites regress: a redesign reintroduces client-side rendering, a CDN rule gets switched on, a redirect chain appears.
We also set up Bing Webmaster Tools, whose AI Performance report shows Copilot citations and the grounding queries the model generated. It is the richest first-party AI data that currently exists and it is free, whether or not you keep working with us.
What we will not tell you
A date by which the answers change. There is no index status to check and no platform publishes its retrieval cycle. We will give you dates for the implementation work, because that is ours to schedule. Treat any specific timeline promise, ours included, as a sales tactic.
Common questions
Usually one to two weeks, and that part we can put dates on because it is our own work. The sampling has to run repeatedly rather than once, which is what takes the time; a single run of each prompt could be done in an afternoon and would tell you nothing you could act on.
No, and quite often stage one is the whole engagement. If the baseline shows your pages render fine, the answering crawlers are unblocked, there are no dead URLs and you are already described accurately on the sources that get cited, then there is nothing for us to fix and we will say so.
Either your developer with our spec, or us. The spec is the same document in both cases: the exact pages, the exact robots.txt and CDN changes, the exact redirect list, and the rewritten opening paragraphs. Most of it is server configuration rather than design work.
Then we find out in stage one rather than three months in. Several hospitality platforms render menus, hours and availability client-side, which makes them invisible to AI crawlers. If that cannot be changed, the mechanical prerequisite fails and we will tell you the retainer is not worth taking.
Start with the baseline
Book a 30 minute call. We will fetch your site the way the crawlers do, read your robots.txt live, and tell you which of the four prerequisites is failing.