Why AI search is ignoring your website

In one paragraph: generative engine optimisation is the work of making a business legible to AI assistants. An assistant cannot click around, cannot run your scripts reliably, and cannot verify a claim it only finds on your own homepage. It reads plain text, cross-checks facts against independent sources, and names the businesses it has the most confidence in. Visibility in AI search is therefore a consistency and readability problem long before it is a content marketing problem.

What actually changed

Traditional search gave the buyer a page of options and left the judging to them. An assistant does the judging first and hands over three or four names. There is no page two, and there is rarely a second opinion. Independent studies through 2025 and 2026 agree on the direction even where they disagree on the exact figures: a growing share of product and supplier discovery now begins inside an assistant, and the people who arrive from one convert far better than ordinary organic visitors, because the assistant has already done the shortlisting. Fewer visits, further along the decision.

The same surveys show most businesses say they intend to optimise for this and far fewer have done anything about it. For a small Irish firm, that gap is the opportunity. This is not an expensive channel to enter. It is mostly hygiene.

The threat is not that AI answers the question. It is that it answers the question without you in it.

Step 1: Audit what the models already say about you

Before changing anything, find out what the assistants currently believe. Open ChatGPT, Claude, Gemini, Perplexity and Grok and run the same three prompts in each, then keep the answers in a document so you can compare them again in a month.

1. The identity prompt. "What does [your business] do, who is it for, and where is it based?" Wrong or vague answers tell you your description is not corroborated anywhere.
2. The buyer prompt. Ask the question a customer would actually ask, in their words: "Who can help a 30-person Dublin practice automate its intake paperwork?" Note whether you appear in the first few names, and who does.
3. The blocker prompt. "Based on what you can read at [your URL], what would stop you recommending this business?" The answers are often blunt and useful.

Answers vary between runs, so do each prompt twice. You are looking for patterns, not verdicts.

Step 2: Make the page readable without JavaScript

Many modern sites build their text in the browser. A human sees a polished page; a crawler that does not execute the scripts sees an empty shell. Check yours in ten seconds: load a page, view source, and search for a sentence you know is on it. If the sentence is not in the source, assume an assistant will never read it.

• Server-render the substance. Titles, headings, body copy, service descriptions, prices where you publish them, and FAQ answers should all be present in the HTML source.
• Stop hiding the good part. Accordions and tabs are fine when the text is still in the source. They are a problem when the content only loads on click.
• Publish the long version. Full case notes, complete method descriptions, transcripts of your talks and videos. Teasers give a model nothing to quote. Substance gives it something to cite.
• Add structured data. Organization, LocalBusiness, Service, Article and FAQPage schema tell a machine what each page is, rather than making it guess.

Step 3: Say exactly the same thing everywhere

Assistants do not take your homepage at its word. They cross-reference structured databases and third-party profiles, and when those sources disagree they hedge, and a hedge usually means naming somebody else. Alignment is the cheapest win available.

• Pick one sentence. One description of what you do, who for, and where. Not a family of similar sentences.
• Repeat it word for word. Website, LinkedIn company and personal pages, Google Business Profile, Crunchbase, industry directories, chamber listings, conference bios, podcast show notes.
• Fix the machine-readable sources. Wikidata entries, where they exist, and any classification fields such as industry, occupation or field of work. These are read directly, not inferred.
• Keep the details identical. Legal name, trading name, address format, phone number, category. Small inconsistencies do real damage here.

When several independent sources state the same fact in the same words, a model will state it too. That is the whole mechanism.

Step 4: Fix the plumbing so the pages can be found at all

No assistant recommends a page it has never seen. Indexing is unglamorous and it is where most new sites quietly fail.

• Submit a clean sitemap to Google Search Console and Bing Webmaster Tools. Bing matters more than it used to, because several assistants lean on its index.
• Check indexing monthly. Search Console will tell you which pages are excluded and why. Ten minutes, once a month.
• Let the AI crawlers in. Read your own robots.txt. Plenty of sites block GPTBot, ClaudeBot, PerplexityBot and Google-Extended by accident, then wonder why they are never quoted.
• One canonical address. Pick www or non-www, redirect the other permanently, and make every canonical tag absolute. Two versions of a site split your evidence in half.
• Consider an llms.txt file. A plain-text summary of your business and key pages. Support is still uneven, but it costs an hour and it forces you to write the one sentence from step three.

Step 5: Write headlines that match how people actually ask

People type differently into an assistant. Not "CRM software Ireland" but "what CRM suits a five-person estate agency that also does lettings?" Assistants also break a question into sub-questions before answering, so a page that answers one sub-question cleanly gets pulled into the answer.

• Collect the real questions: sales calls, support tickets, the wording buyers use in email.
• Make each of those a page or an H2, phrased as the question itself.
• Answer in the first two sentences underneath, in a form that stands alone when quoted.
• Keep one idea per section, so a model can lift it without dragging in the rest.

The uncomfortable part: most citations are not yours

The research on AI citations is consistent on one point: the large majority of what gets cited is third-party material, not the company's own site. That makes sense. A model has no way to verify a self-description, so it prefers a source with nothing to sell.

So the work that moves the needle most is the work you cannot do inside your own CMS. Get into the industry roundups and "best providers for X" listicles. Say yes to podcasts. Write the guest piece for the trade association. Answer real questions in the communities where your buyers complain, without pitching. Publish something a journalist can quote. None of it is fast, and it compounds in a way that on-page work does not.

The one-page checklist

1. Audit. Run the three prompts across five assistants and record the answers.
2. Render. Confirm your real text appears in view-source.
3. Align. One description, word for word, everywhere.
4. Index. Sitemap submitted, robots.txt open, one canonical host.
5. Phrase. Titles and headings shaped like the questions buyers ask.
6. Amplify. Earn independent mentions, patiently.

How Artellis can help

This page is built the way it recommends: text server-rendered, schema in the source, headings shaped as questions, one description of the business repeated everywhere it appears. That is the same job we do for clients.

• An AI visibility audit across ChatGPT, Claude, Gemini, Perplexity and Grok, with the exact prompts, the answers, and who is being named instead of you.
• The technical fixes: rendering, schema, canonicals, sitemaps, robots and an llms.txt, done rather than recommended.
• Identity alignment across your site, LinkedIn, Google Business Profile, directories and machine-readable sources.
• A question-led writing plan built from the questions your buyers actually ask, with a cadence you can keep.

If you would rather see where you stand before committing to anything, book a twenty-minute workflow call and we will run the audit prompts live on your site.