How to improve your AI visibility: start with what AI already reads

“OK, AI doesn’t recommend us. What do we actually do about it?”

The advice on offer is long: add schema, publish an llms.txt file, restructure content into chunks, build authority, write more blog posts. None of it is wrong. But most of it starts from the wrong end.

Since October 2025, Avisible has run about 90,000 AI queries for hotels, gym chains, consumer brands, industrial companies, professional services, congress organisers and non-profits. The companies that had done the most technical work were not the most visible. The fixes that mattered most were usually small, cheap and in plain sight: a fact that wasn’t written down, a page AI already read that left out the product name, two websites that disagreed.

This is the order Avisible recommends:

1. Measure the right thing. Separate the questions that name your company from the questions that do not.
2. Fix what AI already gets wrong. Name, category, contact details, terms.
3. Put the facts in the page text. Not behind a script, not in an image, not in a PDF.
4. Fix the pages AI already reads. They are rarely the ones you would pick.
5. Write the answers buyers ask for. In the buyer’s words, not your category’s.
6. Be present where AI reads. Marketplaces, directories, industry associations, review sites.
7. Measure again. Your 15–20 most important questions, by AI system and by language.

Steps 2–4 take hours, step 5 takes weeks and step 6 takes months. That is why the order runs this way. The reasoning follows.

One technology company Avisible measured had done almost everything on the usual checklist. Every page was indexable, an llms.txt file was live, all AI crawlers were allowed, and the average page had over 900 words. In questions that didn’t name the company, AI still named it very rarely.

Two other sites had 100% schema coverage. Both still had large gaps in open questions. A tourism operator’s site scored 94% on performance, yet AI told travellers its tours cost “from €0”.

Technical work is the floor, not the ceiling. It lets AI read your site. It does not give AI anything worth repeating.

Before fixing anything, split your questions in two: those that name your company and those that don’t. In Avisible’s measurements, almost every brand is named in 90–100% of the first kind. In the second kind, the range is 0–40%. A single blended score hides the whole problem.

Avisible calls that difference the discoverability gap: the distance between how often AI names a company when it is asked about by name, and how often AI names it when a buyer describes a need without naming anyone.

Then ask the questions your buyers ask, in their words and in each language you sell in, and look at three things: who AI names, what words it uses about you, and which sources it cites. The sources are your work list.

The cheapest and fastest gains come from facts AI already repeats, but wrongly. These fixes are usually an afternoon of work.

One company Avisible measured had a Google Business Profile that listed it as a bicycle rental service. AI repeated this word for word. The same brand appeared under six spellings. A marine services firm had two websites that gave different details for four offices, and AI repeated both versions. A consumer brand’s two online stores disagreed on shipping terms, so AI quoted the more expensive ones to US buyers.

The fix is dull and effective: one name, one set of facts, the same on every site, profile and directory. If your company shares a name or an abbreviation with another organisation, add Organization schema with sameAs links. One congress organiser was confused with a technical university in over 30 source links. That is exactly what this markup is for.

AI does not click, scroll or open tabs. If a fact only appears after a script runs, inside an image, or in a PDF, AI often never sees it.

The tourism operator’s prices, durations and FAQ answers all loaded by script. AI answered money questions with “from €0” or sent people to marketplaces. The same site’s cancellation terms were plain text, and AI got them exactly right, every time. A consumer brand put key product information in images without alt text. To AI, that information did not exist.

Avisible’s favourite detail from this dataset: in 16 answers, AI read the tourism operator’s own page and still didn’t name the company, because the company name was not in the body text. Name yourself next to your facts.

This is the step most companies miss, and it often has the best return.

AI already cites some of your pages. They are rarely the ones you would pick. For one industrial manufacturer, the page Google’s AI cited most (in 39 answers) was an application note. It never mentioned the product the company most wanted to sell. Adding the product name to that one page is a ten-minute job.

A congress organiser’s most-cited page was its finance services page, because it had client testimonials. The lesson was not to write more about finance. It was to add testimonials to the other service pages. The marine firm’s most-cited Finnish-language page was one surveyor’s profile that listed the types of cases he had handled. So the recommendation was to give every surveyor a profile like that.

Look at what AI already reads and why. Then do more of that, and fix what those pages leave out.

Only now do you write new content, and it should be specific. In Avisible’s data, the pages that won the answer were application pages, office pages that list services and ports, pages with published prices, and pages that say who the product or service suits.

Write in the buyer’s words, not your category’s. Answer the question in the first two sentences. Put numbers, prices and conditions in plain text. Write about the situations you want to be recommended for, not only the one you are famous for. A brand known only for extreme weather will not be recommended for everyday use until it writes about everyday use.

Also read what your content implies. A tourism operator’s “largest in the Nordics” became “large, commercial, at times touristy” in AI answers. Its job adverts became “guide quality varies by season”. AI repeats everything, not only your marketing.

When a question doesn’t name you, AI reads other people’s sites: marketplaces, city and travel guides, resellers, review sites, Reddit, industry associations and “best of” lists. Each language has its own set.

This is where small fixes also pay off. An industrial manufacturer’s product appeared in 15 Gemini answers as an unnamed reseller listing, because the reseller’s product title left out the model name. A hotel had a sauna page and a local tourism article about its sauna, but AI didn’t connect them to the hotel. Updating the listings on the city tourism sites was one of the hotel’s quick wins.

Decide where you need to be, market by market and language by language, and make sure those sources say the right thing about you.

AI answers change with every model update. Pick 15–20 “North Star” questions that matter most to your business, and track them over time by AI system and by language. Track the competitors AI actually names, not only the ones in your plan.

Most companies start with steps 5 and 6, because content and PR feel like progress. But new content built on wrong facts, hidden prices and a confused name inherits all of those problems.

Steps 2–4 are cheap. In one hotel audit, the report listed eight quick wins that took 30 to 60 minutes each. In Avisible’s Finnish audits, none of the top actions required a site redesign or a developer. Fix what is there, then build.

The findings are based on Avisible audits and Snapshots run between October 2025 and September 2026, about 90,000 AI queries in total, in hospitality, retail, consumer brands and e-commerce, industrial B2B, professional services, congresses and events, and non-profits, in the Nordics, the DACH region, the UK and the US. Each audit asked 100–200 questions, repeated several times, in ChatGPT, Gemini and Google AI Overviews, and most audits also included a technical crawl of the company’s site.

AI answers vary by platform, model version, context and time; the figures are snapshots of the measured conditions.

Read also: Why doesn’t AI search find our company?

Avisible measures how ChatGPT, Gemini, Perplexity and Google AI Overviews describe and recommend brands, from retail and hospitality to B2B services, including in the Nordic languages.

The Visibility Snapshot uses 900 AI queries to show where AI finds your company, which pages and sources it reads, and which fixes come first. Delivered in three working days, €499.