AI discovery

How to improve a recruitment website's chances of appearing in AI answers

How do we improve a recruitment website's chances of appearing in AI answers?

Explain the agency’s specialism, evidence and next step in readable pages, check the search and assistant access you intend to allow, then measure what people and platforms actually do. Server-delivered content, accurate identity and clear links give readers a usable source. Each provider decides whether to find, use or cite it; a technical pass cannot promise an employer enquiry.

01Expertise to enquiryIllustration
Expertise to enquiry: Start with the deskState the sectors, roles and situations your agency understands, with evidence you are allowed to publish. Agency expertise → Readable website → Employer enquiry. This illustrates the method. Rankings, indexing and AI citations depend on external systems and are never guaranteed.
Step 1 of 4
  1. State the sectors, roles and situations your agency understands, with evidence you are allowed to publish.

  2. Deliver the important words in server HTML. Structured data describes those same visible facts.

  3. A person following a referral, a search engine or an AI reader can assess the specialism and how you work.

  4. Route an employer enquiry to the right conversation. Measure it separately from candidate applications.

This illustrates the method. Rankings, indexing and AI citations depend on external systems and are never guaranteed.

The short version

An employer asks an assistant for a specialist agency. The reply may name three firms, cite two websites, mention yours in passing, or mention none of you. Those are different facts, and folding them into one cheerful percentage is how a report becomes impressive while remaining commercially empty.

You can commission a clearer source: the desk’s judgement in readable pages, evidence beside the claim, a sensible route to the right person. You can also check whether the readers you intended to allow can reach it. Those are deliverable changes, while selection in somebody else’s answer remains their decision.

Seven observations, counted separately

GEO and AEO are labels for work around discovery and answers. The useful question for an agency director is what changed, where it was observed, and whether a suitable employer did anything afterwards. A report can collect seven different events under “AI visibility” and leave you unable to answer any of those questions.

Extraction needs its own check too: can a reader recover the correct specialism, price conditions or ATS limitation from the page? An answer that names the agency but invents what it does needs a correction, even if the mention count looks healthy.

Google finds further pages through ordinary HTML anchors with real href values. A library that exists only as a data file, or a method that can be reached only by a scripted click, is not a discovery path. JSON-LD is how we write the visible ScopeSite and RWD facts down, and a format is all it is.

  • Crawlability: a wanted crawler can request the page and follow its ordinary links.
  • Indexing: the URL is in that product’s index, or a report says it is not.
  • Retrieval: observed evidence of material being fetched or used, with unknown internal steps left unknown.
  • Brand mention: the agency or identifiable desk appears in the answer text.
  • Source citation: one of its URLs is shown as a supporting link.
  • Visits: a session reached the public page under the configured analytics rules.
  • Qualified enquiry: a suitable employer started a conversation the agency can actually trace.

Google, ChatGPT, Claude and Perplexity have different controls

Start with the product you intend to reach. Allowing a training crawler is a choice about potential model training, and cannot purchase a place in a search answer. An assistant can combine model knowledge, the conversation, search results and other tools. A model name alone does not tell you which route supplied the answer.

  • Google Search: its AI optimisation guide says ordinary search foundations apply. It does not use llms.txt for Search, require special AI schema or prescribe a fixed answer length. Keep indexing checks separate from browser-agent tests.
  • ChatGPT Search: OpenAI documents OAI-SearchBot for search and GPTBot for potential training. ChatGPT-User handles some user-directed visits. Search and training permissions are independent; a user-directed visit is not proof of search inclusion.
  • Claude: Anthropic distinguishes Claude-SearchBot, Claude-User and ClaudeBot. These cover search, user-directed retrieval and potential training respectively, with documented robots.txt controls for each.
  • Perplexity: PerplexityBot supports search results and Perplexity-User handles user-requested visits. Its documentation describes neither as a foundation-model training crawler, and says user-requested fetching generally ignores robots.txt.
  • Bing and Copilot: Bing Webmaster Tools reports citations across supported Microsoft surfaces and a sample of grounding queries. Those figures do not establish position within an answer or coverage of every assistant.

Why browser agents and llms.txt deserve a separate check

A browser agent may use the rendered screen, HTML and accessibility information to complete a task. Google’s web.dev guidance therefore calls for clear controls and stable layouts. For a recruitment site, test whether the employer can find the relevant service and enquiry route, with the buttons and labels making sense, before anyone celebrates the schema.

The llms.txt proposal provides a curated map for agents, and Lighthouse has an agentic browsing audit for it. Chrome’s documentation currently treats a missing file as optional. A valid, useful map is worth maintaining when you provide one; it is neither an access rule nor evidence that an assistant selected the business.

That sits alongside Google Search’s separate statement that llms.txt does not help or harm its search visibility. The tool check and the Search guidance answer different questions. Check the actual file, its links and its agreement with the public pages, then record the result under resource quality.

A dated starting count, with the limits still attached

On 25 September 2026 we pulled Recruitment Web Design’s own Search Console web-search data for 26 August to 22 September: nine page rows, eighty-two page-level impressions and two clicks. The website-company comparison article accounted for 18 impressions and one of those clicks, pricing had eight impressions and the other click, and the website-design, ATS and software pages had impressions and no clicks.

That is a starting count for this site, too small for a before-and-after study, a ranking or a speech about momentum. The useful part is the discipline: a date range, a property, a page list, and a refusal to pad the story.

The same window does not tell you what an assistant said. A mention or citation check needs its own log: write the buyer question first, then the engine, the mode, the date and the geography, and score the seven fields independently. If you skip the prompt and the date, you do not have a measurement. You have a memory of being pleased.

No assistant transcript is invented here, because inventing one would be a rather neat way of demonstrating the problem.

The approved named proof of the wider method sits on the published H4TLT case study, which is about workplace hearing surveillance. Read that page for the evidence the client has permitted, which is not a recruitment placement result and does not travel into this site’s two clicks.

The work you can actually commission

Start with the conversation the director already has when the work is going well: which roles, which employers, and what “good” looks like once you drop the slogan. If that explanation lives only in someone’s head, or in a PDF last saved under a consultant who has since left, the website is asking a stranger to finish the brief.

Write the missing facts on the relevant page, in the first HTML response. Google can run JavaScript, while other readers may never execute a script, so putting the important words in the first response is a design choice with a reason rather than a claim that every bot is helpless or that inclusion has been decided.

Then check who is allowed in. A firewall challenge can stop the crawler you meant to welcome. Match requests against the provider’s current verification guidance, including published IP ranges where available; a user-agent name can be copied. OpenAI says a robots change can take around a day to apply, so an immediate retest is only an immediate retest.

Keep the names consistent: SCOPESITE LTD, trading as ScopeSite Digital Studios, supplies the Recruitment Web Design brand, and Dan Cartwright is the author people can look up. Structured data should repeat visible facts without inventing an office, a rating or a second company. A graph connects identities; it does not appoint the business to an assistant’s shortlist.

  • One employer question per page, answered near the top.
  • Approved evidence, with a sentence about what it does not prove.
  • A next step an employer can use without uploading a CV.
  • Crawler rules checked against the bots you intended, rather than a generic “block AI” preset.
  • A written prompt set for later comparison, instead of a new article for every wording of the same question.

What the agency supplies, and what sits outside the build

ScopeSite can structure, deliver and test the public explanation. The desk still has to supply the judgement. If the specialism is unresolved, the website cannot make it true, and publishing a planned service as an established one is how you earn a very confident wrong answer later.

Vacancies, if they form part of the public story, follow the system that already owns them. A feed that is a day behind will give search and assistants a day-behind story. That is an operations problem with a website symptom.

  • The agency supplies: specialism, geography, approved proof, real employer questions, access, and a person who can sign off wording.
  • RWD work, delivered by ScopeSite, controls: page structure, server-delivered HTML, ordinary crawlable links, matching structured data, crawler-policy checks, and the seven-column log.
  • Platforms control: crawling, indexing, retrieval, citations and whatever they decide to say on Tuesday.

How the work is done

Discovery starts as a page review. We take one live URL, one employer type and the questions that person would actually ask, then record what the first screen explains, what evidence is visible, where the enquiry goes, and what a first-response reader receives before scripts run.

The change is then small enough to describe: rewrite the unclear specialism, move the answer up the page, fix a blocked crawler, align the organisation name, and stop a closed vacancy from volunteering for work. We do not commission a hundred near-identical location pages because a keyword tool looked hungry.

Use two distinct tests. Give a reader the page and ask factual questions to check extraction, including the limits on prices or integrations. Then ask the buyer question in the actual search or assistant product without supplying the website. A local model or embedding test can help find missing evidence, but success with a supplied page does not show that ChatGPT, Google or another product would discover it unaided.

Write the questions before testing, including unbranded supplier choice, costs, ATS retention, evidence and the next step. Record date, product, mode, country and personalisation, save the answer and cited URLs, and keep failures. Rephrasing until your name appears produces a particularly agreeable spreadsheet.

Use comparable observation windows after a recorded change. Keep index status, provider reports, consent-based visits and qualified enquiries in separate records. Repeat the same prompts, with fresh context where possible, and report unavailable products as untested. A single answer is a snapshot; a before-and-after difference still needs other explanations considered.

What this cannot promise

This work cannot make an assistant recommend you, stop a competitor with a clearer page from being named first, turn 82 impressions into a market, treat H4TLT’s permitted case study as a recruitment testimonial, or use another supplier’s article as a draft.

Other recruitment firms already publish GEO and AEO guidance. That is useful, because it saves us from a silly uniqueness claim. The difference worth offering is a method you can inspect: retained ATS where the connector allows it, labelled demonstrations, dated sources, and counts that refuse to merge.

Give it something true to read, then count what happened.

Recruitment Web Design’s own Search Console window, 26 August to 22 September 2026

Authenticated web-search data retrieved on 25 September 2026: nine page rows, 82 page-level impressions and two clicks. Used as a dated starting count for this site, with no claim that the library improved anything.

  • The website-company comparison article recorded 18 impressions and one click, pricing recorded eight impressions and one click, and the website-design service, ATS integrations and recruitment software pages recorded 12, 12 and 11 impressions with no clicks.
  • The sample is too small for a trend, a ranking claim or a commercial conclusion. That limit is part of the example.
  • The same retrieval is web search, and it does not say what ChatGPT, Claude, Perplexity or Grok answered on those dates.
  • A separate answer log would use a written buyer prompt, the engine, the mode, the date and the seven independent fields, with no invented assistant answer attached.
  • The approved named method evidence remains the published H4TLT case study in occupational hearing surveillance. Read that page for the client’s own evidence, which is not a recruitment placement result.

What we count, separately

  • Crawlability: A wanted crawler can request the page and follow ordinary HTML anchors with real href values, unless robots, firewall or authentication choices block it. Training-crawler permission is a separate decision.
  • Indexing: The property’s search index contains the URL, or a Search Console report says it does not. Indexing is not guaranteed after a crawl, a sitemap row or an IndexNow receipt.
  • Retrieval: Record observable retrieval evidence, such as a provider’s grounding report or a verified page request. A citation alone does not reveal when the page was fetched; the full internal process may remain unknown.
  • Brand mention: The agency name, Recruitment Web Design brand or identifiable desk appears in the answer text for a written prompt, on a recorded date, engine and mode. A mention can be accurate or wrong.
  • Source citation: A URL from the agency, or another agreed source page, is shown as a supporting link. Google’s generative AI report, where available, records link impressions. Bing reports citations across supported surfaces and a sample of grounding queries.
  • Visits: A session reached a public page under the configured analytics rules, recorded in its own visit column beside mention, citation and enquiry rather than standing in for them.
  • Qualified enquiry: A suitable employer started a conversation that the agency can tie to a discovery path with its own records. Planner use and demo clicks stay in their own columns.

What Recruitment Web Design controls

  • The public wording, page structure and first HTML response we are commissioned to deliver.
  • Ordinary crawlable anchors between the pages we publish.
  • robots and firewall recommendations for search crawlers, kept separate from training-crawler decisions.
  • Consistent ScopeSite organisation, RWD brand and author names that match the visible page.
  • Structured data that restates those visible facts, with no extra invented company, office or rating.
  • The written prompt set, observation dates and the seven-column log.

What the agency supplies

  • The specialism, geography and employer type the agency wants to be found for.
  • Approved evidence and the sentence that limits what it proves.
  • Website, CMS, DNS and approval access.
  • Whether live vacancies form part of the public story, and which system owns them.
  • The employer questions already asked on calls.

What this cannot guarantee

  • Crawling, indexing, retrieval, appearance, citation or recommendation.
  • That a prompt will trigger a live web fetch.
  • That schema depth, llms.txt or a readiness score will cause an assistant to name the agency.
  • That visits become qualified enquiries.
  • A first-or-only position among recruitment suppliers.

Sources

  1. Optimizing your website for generative AI features on Google Search

    Google Search Central. Checked . Primary technical source. Current Google rules for useful original pages, ordinary SEO, and the absence of special AI schema or llms.txt requirements. Google Search guidance only. It does not describe ChatGPT, Claude, Perplexity or Bing, and it does not guarantee indexing or appearance.

  2. Understand JavaScript SEO basics

    Google Search Central. Checked . Primary technical source. Confirms that Google can render JavaScript, so JS-blind claims cannot be applied to Google Search. Does not describe every AI crawler, and rendering is not a ranking or citation promise.

  3. Structured data general guidelines

    Google Search Central. Checked . Primary technical source. Markup must match visible content. Valid JSON-LD does not guarantee a rich result. Policy for Google rich results, not a mechanism that causes AI recommendations.

  4. Link best practices for Google

    Google Search Central. Checked . Primary technical source. Library and service links must be ordinary anchors with real href destinations and descriptive text. A method page reachable only by script is not a discovery path. Google Search crawl discovery only. It does not describe assistant retrieval or guarantee indexing.

  5. Generative AI performance report

    Google Search Console Help. Checked . Primary technical source. Defines AI Overview and AI Mode link impressions as a Search Console observation. Impressions are not recommendations, citations in other assistants, or enquiries.

  6. Overview of OpenAI crawlers

    OpenAI. Checked . Primary technical source. Separates OAI-SearchBot, GPTBot and ChatGPT-User. Training permission is not required for ChatGPT Search. OpenAI product controls only. robots.txt updates can take around 24 hours to apply.

  7. Anthropic crawlers and site-owner controls

    Anthropic. Checked . Primary technical source. Separates ClaudeBot for potential training, Claude-SearchBot for search and Claude-User for user-directed retrieval. Access controls describe those bots. They do not guarantee a Claude answer or describe every browser-agent action.

  8. Perplexity crawlers

    Perplexity. Checked . Primary technical source. Distinguishes search indexing from user-directed fetching and recommends verifying published IP ranges alongside user-agent names. Neither bot is described as foundation-model training. A permitted request does not establish a recommendation.

  9. Build agent-friendly websites

    Google web.dev. Checked . Primary technical source. Explains browser agents using screenshots, HTML and accessibility information, with clear controls and stable layouts. Task usability guidance. It does not establish Google Search ranking or citation behaviour.

  10. Lighthouse agentic browsing: llms.txt audit

    Chrome for Developers. Checked . Primary technical source. Documents an optional llms.txt audit in the agentic browsing category, with a missing file currently treated as not applicable. An audit result describes the tested resource. It does not measure search rankings, citations or enquiries.

  11. The llms.txt proposal, version 2

    Jeremy Howard, llms-txt. Checked . Primary technical source. A proposed curated map for agents, with concise context and links to useful content. A proposal rather than a universal provider requirement, access-control mechanism or ranking guarantee.

  12. JSON-LD 1.1

    W3C. Checked . Primary technical source. Named graphs and identifiers for connecting real entities. A format choice, not inclusion in any vendor knowledge graph.

  13. AI Performance in Bing Webmaster Tools

    Microsoft Bing. Checked . Primary technical source. Defines citations across supported Microsoft surfaces and a sample of grounding queries. The grounding-query sample is incomplete. Citation totals do not measure ranking, all-AI share or enquiries.

  14. Spam policies for Google web search

    Google Search Central. Checked . Primary technical source. Blocks scaled unoriginal pages and copied competitor articles as a delivery method. A policy boundary, not a traffic forecast.

  15. H4TLT published case study

    ScopeSite. Checked . Approved claim. The only currently named client proof. Method evidence in occupational hearing surveillance, not a recruitment result. Do not restate unverified traffic or revenue, or treat a cross-sector case as a recruitment outcome.

  16. Recruitment Web Design Search Console web-search pages, 26 August to 22 September 2026

    Google Search Console, retrieved for the RWD library plan. Checked . First-party measurement. Dated first-party starting count: nine page rows, 82 impressions and two clicks. Used as a measurement example, not a campaign result. A short, low-volume window. It cannot support trend, ranking or commercial conclusions, and it is web search rather than every assistant.

Related pages

Show us the bit that is getting in the way

Pick a website or desk task. We’ll establish what is feasible.