Updated 16 October 2026. This article replaces earlier claims about a six-week client result, universal AI rankings and scanner capabilities. Those claims should not be used as evidence or as a promise of future results.
A recruitment agency can have a useful website and still be absent when an employer asks ChatGPT for a supplier. That answer deserves investigation. It does not, by itself, tell you whether the problem is the website, the question, the sources selected or the assistant’s particular response.
Start with a question you can test, then keep the evidence attached. A website score, an assistant mentioning your agency and an employer submitting an enquiry are different observations. Putting them in the same green circle saves dashboard space, but makes the business decision rather harder.
Choose the employer question first
Write down what a prospective client might need: a recruitment partner for a particular role, sector or location; evidence that the agency understands the work; or an explanation of how an assignment starts. Use the language that appears in real enquiries where you have permission to use it.
Keep a small, fixed set of questions. Include questions about your named agency and questions that do not mention it. “What does this agency do?” checks a different problem from “Which agency should we consider?” Record the wording before testing, so a disappointing answer does not become a different question.
Make the relevant page useful
Read the page an employer should find. Can they identify the service, the roles covered, the area served, the supplier and the next step? Does the evidence support the claim being made? A list of sectors is less useful than a clear account of the work the agency actually undertakes.
Keep names and relationships consistent across the site. Explain a specialist service’s relationship to its parent company. Publish genuine experience and approved client evidence, with enough context to understand it. A company does not acquire another office because somebody has added a postcode to its structured data.
Check access without treating AI as one visitor
Inspect the HTML returned by the server as well as the finished browser page. If important service information appears only after scripts run, some retrieval tools may miss it. Google Search can process JavaScript, while browser agents may inspect a rendered page. Test the relevant route before prescribing a rebuild.
Crawler controls also have different jobs. OpenAI distinguishes search, potential training and user-requested fetching. Anthropic documents separate crawler roles too. Check the actual user agent, robots rules, response status and any verified firewall evidence. Permission to crawl does not establish that an assistant selected or cited the page.
Use structured data to describe supported facts
Relevant schema can describe the organisation, services, articles and other information visible on a page. Validate the syntax and check the facts. If the page changes, the markup needs checking too.
Google says structured data is not required for its generative search features. It remains useful for eligible rich results, but there is no special schema recipe that guarantees an AI recommendation. Google’s Rich Results Test checks its supported features; an empty result does not establish that a page contains no structured data at all.
An optional llms.txt file can provide a curated map for tools that choose to read it. Keep it accurate if you supply one. It does not replace public content, access controls or the separate work of measuring actual answers.
Record what the assistant actually returned
For each question, save the product, mode, date, relevant location or personalisation settings, answer and cited URLs. Note whether a search visibly ran. Use a fresh context where practical and record any access failure. A consumer product and its provider’s API can produce different observations.
Score the parts separately: whether the agency was mentioned, whether it was described correctly, whether a useful source was cited and whether the answer gave a sensible next step. One favourable screenshot cannot establish a permanent position, and repeating a question until the agency appears makes the record less useful.
Run a separate test with the page supplied explicitly. That can reveal missing or ambiguous information. It does not show that the assistant would have discovered the page unaided.
Repair the fault you can demonstrate
A missing service description needs a content repair. A failed request needs an access investigation. Contradictory company names need reconciliation. Keep the work connected to the finding, then repeat the same check after publication.
Watch search data and consent-based website analytics alongside the answer records. An assistant citation, a referred visit and a qualified enquiry each need their own evidence. Agree a review schedule, rather than promising that every business will appear in answers after a fixed number of weeks.
Where ScopeSite’s work fits
Our published H4TLT case study sets out the approved evidence for work with Hear 4 The Long Term, a workplace hearing-surveillance business. Read its scope directly; it is not a recruitment result or a timetable for another company.
CAFMO is live at canaifindme.online. Can AI Find Me Online is ScopeSite’s application for page checks, model answer evidence and a saved improvement plan, with monthly subscription plans. V.O.I.C.E. remains the methodology and engine name. The earlier scanner keeps separate accounts and credits; purchased legacy credits do not expire or automatically become a subscription. Model API observations can differ from the consumer assistant applications.
For recruitment work, the Recruitment Web Design discovery guide sets out the checks in more detail. Begin with the employer question and the page meant to answer it, then decide what the evidence justifies changing.


