How does your company get into AI recommendations?

Three engines, three different games

Almost every guide on the Hungarian market treats “AI” as one thing. AVE Studio's measurement says that is the most expensive misunderstanding in the field: the three major engines work so differently that identical effort delivers everything on one and nothing on another.

The gap is sharpest on “who should I avoid” questions: there, ChatGPT did not search once in 30 runs in AVE Studio's measurement. So when your buyer asks who to steer clear of, ChatGPT works purely from what it learned earlier — a fresh page cannot reach that layer, only what others write about you, over a long enough period and in enough places.

Ten steps, in this order

The order is not cosmetic. Without the first three the rest cannot take effect, and from the eighth onward you can measure whether any of it did.

1. 1. Check whether machines can reach you at all

   Three separate classes of bot live in robots.txt today, and most companies know about only one. There is the search indexer (Googlebot, Bingbot), the retrieval-time reader (OAI-SearchBot, PerplexityBot) and the trainer (GPTBot, ClaudeBot, Google-Extended). Each can be allowed or blocked independently, and most absences are decided here rather than in the content. A case we measure often: a trade publication lets the search crawler in but blocks the training bots — your article there stays indexable, yet can never enter ChatGPT's memory. On your own site the decision is yours; when you publish elsewhere, check first.

   Concretely: open your-domain.com/robots.txt and look up GPTBot, ClaudeBot, Google-Extended, OAI-SearchBot and PerplexityBot separately. If none of them appear, the “*” rule governs them — and it is probably not what you would have chosen.

2. 2. One question, one page — and let the title be the question

   The pages sitting behind the answers in our measurement have titles that are literally the question a user asks. Not a keyword, not a slogan: a question. It follows that one page should carry one question. If you want to answer five things, you need five pages, not one long piece with five chapters. This is also the cheapest mistake to fix: the title of your existing service page can be rewritten into what your buyer actually asks.

   Concretely: write down the five sentences your buyer actually types. Check which of them you have a page for. The one you do not is the title of your next page.

3. 3. Make the first paragraph a liftable answer

   Engines lift sentences, not pages. If your opening paragraph sets the scene, there is nothing in it to lift. Write it so that on its own, without any context, it fully answers the question in the title — and so that a reader who reads only that paragraph still gets something usable. The rest of the article can then proceed as normal.

   Concretely: read your own opening paragraph as if it were the single sentence a machine will quote. If it is not an answer, rewrite it.

4. 4. Put your name and the claim in the same sentence

   This step comes from a failure of our own. When AVE Studio published its Hungarian model-mindshare study, Perplexity cited its number in five runs out of five — and never once named us. The number travelled; the brand did not travel with it. Since then the rule here is simple: if a sentence contains a number, that same sentence contains the name. Not because it reads better, but because lifting happens at sentence level, and a company name in the neighbouring sentence does not travel with the data.

   Concretely: find every sentence on your site that contains a number and check whether the company name is in it. If not, put it there.

5. 5. Be concrete: numbers, dates, method

   There is nothing to lift from a generality. The strongest Hungarian pages in our measurement are all dense with figures — counts of checkpoints, weightings, scales, numbers of companies analysed. Note that this is about liftability, not proof: a concrete sentence has edges, a general one does not. If you have your own data, publish it together with the method (what you measured, when, how many times). If you do not, an observation from your own work is still more concrete than an industry platitude.

   Concretely: attach a number, a date or a method to each of your claims. If you cannot attach any of the three, the claim probably is not yours.

6. 6. One name, one description — the same everywhere

   If your company is called three different things on your website, on LinkedIn and in the company register, the model ends up with three half-known entities rather than one known one. The same goes for the description: if you say something different about what you do on every surface, what the model learns is uncertainty. Fix the name, the short description, the address and the services in one place and feed every surface from there — structured data (schema.org) included.

   Concretely: write the legal name, a 25-word description, the address and the service list into one file, and copy that to every surface — including the schema.org Organization block.

7. 7. Be where the engine already goes

   In AVE Studio's measurement, linkedin.com sat among the sources in seven of the twelve questions — more questions than any Hungarian company site. The LinkedIn items being cited are not short posts but long, structured articles with question titles. Alongside them, YouTube (six videos) and professional forums appear regularly. In the short run this is worth more than another subpage on your own domain: your own site is the long-term asset, but third-party surfaces are already inside the answers today.

   Concretely: take the one question you answer best and write it up at length on an external surface too — from a different angle than your own page, with a link back.

8. 8. Have a page for the negative question too

   Your buyer does not only ask who to choose, but also who to avoid. Across 178 mention records in AVE Studio's measurement, the engines never once named a company to avoid — they listed criteria instead. So the “who should I avoid” answer gets assembled from a checklist, and it reasons with the framework of whoever wrote that checklist. Publishing honestly when someone should not trust a provider therefore hands the engine your own yardstick — including the part where that yardstick can be turned back on you.

   Concretely: write down the signals that should make someone reject you as well. If that stings, you are probing the right place.

9. 9. Publish the date, and actually maintain it

   Almost every page in our measurement states its publication or update date; the one that does not is the weakest of the set. This is not a trick: without a date neither a reader nor a machine can judge whether the page still holds. If you update it, say so, and let it show in the structured data as well.

   Concretely: put the date in the visible page and in the structured data, and set yourself a quarterly reminder.

10. 10. Measure — do not take our word for it, or your own

    The nine steps above follow from checkable claims, but on your market with your questions things may land differently. The only way to find out is to ask the same questions repeatedly, on several engines, and watch what changes among the sources. A single query proves nothing: answers reshuffle from run to run. Repetition is what turns them into data.

    Concretely: ask the same question on five separate occasions and note which pages appear beneath the answer. The repeat visitors are the real field.

Which lever moves what

The classic lever is the one you know from search optimisation: accessibility, content, structure, freshness. The behavioural lever acts on the model's memory: what others write about you, how consistently, and for how long. AVE Studio's measurement says the two do not move the same engine.

The practical conclusion: if you need a result by tomorrow, do the Perplexity-side work. If the question is why ChatGPT never mentions you, that takes months — and it is not decided on your own website.

Seven questions about yourself — two minutes

This is not a scored quiz but the seven most expensive gaps. Every “no” is a concrete task from the list above.

1. Can you say from memory which bot classes your robots.txt lets in?

   If not, step 1 is yours — and an old setting may be excluding half of what you want.

2. Do you have a page whose title is literally your buyer's question?

   If not, everyone else in the field does, and the engine matches their titles instead.

3. Does your opening paragraph answer the title's question on its own?

   If it merely sets the scene, there is no liftable sentence in it, and the machine lifts your neighbour's.

4. Do the sentences containing your numbers also contain your company name?

   If not, your data is quotable — without you. We measured this on ourselves.

5. Are your name and description identical on your site, on LinkedIn and in the company register?

   If not, the model holds several half-known versions of you, and they weaken one another.

6. Does your company appear anywhere other than your own domain?

   If not, ChatGPT's memory layer has nothing to hold on to — that layer is built from other people's pages.

7. Do your professional pages carry a visible date?

   If not, neither a reader nor a machine can tell whether what you wrote still holds.

Five mistakes we see most often

Frequently asked questions

How long before my company shows up in AI answers?

It depends on the engine, and we will not give you a date. With Perplexity days can be enough: our own model-mindshare study was among its sources five days after publication. ChatGPT's memory layer is considerably slower, and it is not decided on your own site. Anyone quoting a firmer deadline is asserting something they cannot know.

Is ranking well on Google enough?

It does not hurt, but it is not enough. In AVE Studio's measurement ChatGPT drew on no live source at all in 88% of runs — your Google position cannot act there. With Perplexity, by contrast, fresh and reachable content counts directly. That is why there is no single-sentence answer to how you are doing “in AI”.

Do I need special schema or an llms.txt for ChatGPT?

Structured data: yes, it helps, because it makes your entity unambiguous. The usefulness of llms.txt is not demonstrated today, and it certainly does not replace a sitemap or proper crawlability. If you have to choose, first fix what the crawlers documentably use.

Why isn't publishing more pages enough?

Because volume is not the signal engines respond to. We published two pages of our own in early August, and neither entered the sources in the following measurement — while our months-older study, the one carrying measured data, was cited by two engines. The difference was not the page count but whether the page contained something concrete enough to lift.

What does it cost?

The entry point is the Citation Tracker, from €240 net; one-off instruments run from €240 to €490 net, and continuous measurement from €200 net per month. There are no bundles; we price per module.

Do you guarantee that I will appear in the answers?

No — and if someone guarantees it, ask them what they measure. Neither we nor anyone else can govern what a model returns. Our job is to measure where you stand, tell you what moves it in your specific case, and then measure again to see whether it moved. If it did not, we will say so.

What exactly do you measure?

Not the wording of the answer but the source set: which pages the engine relies on when asked within your category. Every question runs repeatedly, across several engines, and each positive question is paired with its inverse (“who should I avoid”). That is how you find which sources write both the recommending and the rejecting answer — the real decision-makers in your category.

What if the model knows me, but wrongly?

That is more common than being absent entirely, and it is separate work. Models rarely invent from nothing; typically they confuse what they half-know — a similarly named company, say, or an outdated profile. The fix is the same consistency work as in step 6: one name, one description, everywhere, in enough places that the half-known version loses.

Related pages

• When not to trust a GEO agency — seven red flags
• There are two kinds of GEO — classical and behavioural
• Our measurement methodology, in detail
• What a GEO agency actually does