GEO: how ChatGPT, Perplexity and Google AI decide to recommend you

Your customer no longer searches. They ask.

A few years ago, someone who needed a garage typed „car service Chișinău” into Google and got a list of links. Today, more and more often, that same person opens ChatGPT and writes: „where can I get an automatic gearbox repaired in Chișinău at a reasonable price?”

The assistant gives two or three names and sometimes a link. Every other business in the city simply is not in that conversation. They were not rejected — they were never considered.

The scale is no longer small. On 27 February 2026, OpenAI announced that ChatGPT had passed 900 million weekly active users, up from 400 million a year earlier. Not all of them ask about local businesses. But enough do to matter.

At the same time, the habit of clicking is thinning out. A SparkToro study using Similarweb clickstream data, covering January to April 2026, found that 68.01% of Google searches in the United States end without a single click through to a website — up from 60.45% in 2024. The authors state their own limitation: the analysis excludes searches made inside the Google mobile app, where they expect the share to be even higher.

That figure is American, not Moldovan. No equivalent measurement has been published for the Republic of Moldova, and anyone presenting this percentage as „Moldovan” is inventing it. The direction, though, is the same everywhere: the answer reaches the person without the person reaching you.

GEO (Generative Engine Optimization) deals with exactly this: why one business appears in that answer and another does not.

Three surfaces. They are not the same thing.

People talk about „AI” as if it were one place. There are at least three, with different mechanics.

The practical consequence: there is no single fix. A site can appear in AI Overviews and be entirely absent from ChatGPT, because being indexed by Google and being reachable by OpenAI's crawler are two independent things.

How an AI assistant picks who to recommend

A language model does not hold a list of firms in Chișinău that it consults. Simplified, it does three things:

1. Searches the web, when the question needs current or local information.
2. Reads the pages it finds, if it is allowed to read them.
3. Picks the information that looks clearest, most consistent and easiest to summarise.

That gives you three things to fix on your site: access, consistency and quotability. In that order, because each one depends on the one before it.

Step 1: check whether AI crawlers can read your site

This is the most common problem and also the cheapest to fix.

Open your-site.md/robots.txt in a browser. If you see the name of an AI crawler followed by Disallow: /, that crawler is not reading your site.

Who the crawlers are and what each one does

Here is the nuance most guides miss. Not all AI crawlers do the same job.

The difference matters a great deal. Training crawlers take your text to make the model better in general. Search crawlers come for your page now, to quote it in an answer going to a person who is asking at this very moment.

If you want to be quoted, the search crawlers need access. You can block training and stay quotable. These are two separate decisions, and most sites treat them as one.

One more clarification, because it gets confused often: Google-Extended has nothing to do with your position in Google. Blocking it does not remove your site from search results.

How you get blocked without deciding to

This is the part business owners do not suspect.

On 1 July 2025, Cloudflare announced a feature called managed robots.txt. If you enable it, Cloudflare automatically updates your robots.txt file — or creates one if it does not exist — adding directives asking AI bot operators not to use your content for model training. The directives update themselves, with no action from you.

The feature is not applied to everyone automatically: Cloudflare says it is opt-in and that the company merely encourages new sites to enable it. But if someone once ticked that box — you, your developer, or whoever moved the site onto Cloudflare — your robots.txt now contains rules nobody on your team wrote.

It is not the only mechanism. Site themes, security plugins and hosting panels add their own defaults too. The result is the same: a file nobody has ever read decides whether you exist for AI assistants.

What to do

1. Open robots.txt and read it to the end. It takes a minute.
2. Note which crawlers are blocked and ask yourself whether the block was a decision or an accident.
3. If you use Cloudflare, check in the dashboard whether managed robots.txt is on.
4. If you want to be quotable, let the search crawlers through.
5. Check again in a month. Crawler names change and companies add new ones. The official documentation from OpenAI, Anthropic, Perplexity and Google is the only reliable source for the current list.

How widespread is blocking? It depends who you ask.

Worth saying, because you will find wildly different figures. Studies published in 2026 report the share of sites blocking GPTBot at anywhere between roughly 10% and 30%, depending on the sample: some measure a few thousand randomly chosen sites, others only the few thousand highest-traffic sites in the world.

The gap does not mean somebody is lying. It means „websites” are not a uniform group, and large publishers block far more often than a local business does. For you, the global figure does not matter at all. What matters is what your own file says.

Step 2: the same business details everywhere

A model that finds three different phone numbers for your business has a problem. It does not know which one to give, so it often gives none and moves on to the next firm.

Check that these match, character for character:

The places that count: your site, your Google Business profile, Facebook and Instagram, local directories, and any catalogue you were once listed in and have forgotten about.

Businesses that have moved or changed their number in the past few years almost certainly have a problem here.

Step 3: write so you can be quoted

A model prefers text it can summarise without getting it wrong.

Answer directly, at the start. If the page is called „How much does a service cost”, the first paragraph must contain an answer, not the history of the workshop.

Be concrete. Compare:

Put the source next to the number. A model that finds a figure with no source treats it with suspicion. So does a person.

Use the same terms throughout. If one page says „automatic gearbox” and another says „automatic transmission”, you have split the signal about one service in two.

Structure it. Clear headings, lists, tables. A ten-line paragraph is hard to quote.

What this means in a market that searches in two languages

The Republic of Moldova has a quirk that changes the calculation.

According to StatCounter, search in the Republic of Moldova broke down like this in August 2026:

Google dominates, but Yandex is not zero, and those 7.75% are not spread evenly: they are concentrated in the segment that searches in Russian.

For AI assistants, the consequence is not that you should optimise for Yandex. It is that questions about your business get asked in two languages, and the answers differ. An assistant asked in Romanian finds different sources than the same assistant asked in Russian, because the web it draws on is different.

If you only check in one language, you are checking half the market.

How to find out where you stand

Do not assume. Check.

1. Write ten questions a real customer would ask.
2. Put each question to ChatGPT, to Perplexity and to Google.
3. Ask them in Russian too.
4. Screenshot each one with the date visible.
5. Record not only whether you appear, but which sources the assistant cites. That is where the useful information is: if they all cite the same local catalogue, you know where you need to be.

Keep the record in a simple table:

After three months you have a data series. Before that you have impressions.

What GEO does not do

GEO does not guarantee you a place in any answer. Nobody controls what a model says, and answers change from week to week and from user to user.

GEO does not replace SEO. If your site is slow, blocked or empty, there is nothing for a crawler to take from it.

GEO has no payment button. There is no auction where you buy a place in the answer.

And GEO does not come with a score. Anyone showing you an „AI visibility score” from 0 to 100 is showing you a number they invented, not one reported by OpenAI or Google.

What GEO does is put you in a position to be chosen: accessible, consistent and easy to quote. The rest remains the model's decision.

The first thing you can do today

Open your-site.md/robots.txt. It takes a minute. If you find an AI crawler blocked there that you did not choose to block, you have already found the reason you appear nowhere.

Sources

• OpenAI, 900 million weekly active users — public announcement, 27 February 2026.
• SparkToro, „Zero-click searches” — analysis of Similarweb clickstream data, US Google searches, January–April 2026. The study excludes the Google mobile app.
• Cloudflare, „Control content use for AI training” — official Cloudflare blog, 1 July 2025, the managed robots.txt feature.
• StatCounter Global Stats — search engine share, Republic of Moldova, August 2026. Estimates from measured traffic, not a census.
• Official documentation from OpenAI, Anthropic, Perplexity and Google — for the current list of crawlers and supported directives.

Unless stated otherwise, the figures above are measured by third parties, in markets that are not the Republic of Moldova. They are not results we obtained for clients. When we publish our own numbers, they will come with dated screenshots from client accounts.