Artificial intelligence

Why doesn’t ChatGPT mention your business?

Open ChatGPT and ask it what your customer asks: who provides this service, in your wilaya (province). If it does not name you, look first for what stops the assistant from understanding you, or from reaching you at all, before blaming your content. This guide takes apart six causes, starting with the one that cancels all the others, and gives you, for each, the check you can run yourself without buying a tool.

The short answer for a business owner

A search engine shows a list; an answer engine writes one answer. There is no second place in an answer.
Check blocking before content: a shared host can return 403 to AI crawlers without the site owner knowing — ours did until 12 September 2026.
What is not structured text does not enter the answer: a price inside an image, a service inside a PDF, do not exist for the assistant.
Contradictory information is worse than missing information: one price on the site and another on the Facebook page pushes the model to avoid you.
Start by measuring, not by fixing: ask the assistants today and archive their answers — that is your baseline.

The six causes, and the check you can run yourself

Every row here takes minutes to verify and needs no paid tool. Start at the top: the first two causes cancel everything below them, because polishing the copy is pointless while the server keeps the door shut.

Cause	How to check it yourself	What it means	First fix
The server blocks the crawlers	Request /robots.txt with the GPTBot signature, then with a browser one, and compare.	The assistant reaches none of your pages, whatever they contain.	Exclude the crawler signatures from the host’s bot protection.
No structured identity	Open your page source and look for ld+json.	The assistant does not know you are a registered company, where, or for what.	Organization data: legal name, registration numbers, address.
Prices and services inside images	Try to select and copy a price. If you cannot, it is an image.	The question "what does it cost?" finds nothing quotable on your site.	Write the price as text, with what it includes and what it does not.
Contradictory information	Compare name, address, hours and prices across site, page and Google Maps.	The model favours consistent sources and avoids contradictory ones.	Make those four facts identical everywhere they appear.
No direct answers	Do you have a page that answers the buyer’s question in its first sentence?	The assistant finds no quotable sentence and takes one from somebody else.	FAQs worded the way the buyer actually asks, not the way you prefer to say it.
No identifiable author	Does your content carry an author name verifiable outside your own site?	A source with no identity weighs less than one with a face and a track record.	A named author, their experience, and their official profiles linked together.

A search engine shows; an answer engine decides

On Google you get a list: ten links, ads, a map, related questions. You are the one who chooses. Fifth place therefore still exists, and whoever did not win you with a title may win you with a description.

When you ask an assistant you get a paragraph. No list, no ranks. It cites one source or two; everything else does not exist for the person who asked. The difference is not of degree but of kind: in search you compete for a position, in an answer you compete to be mentioned at all.

This is why treating it as a ranking problem is a mistake. The question is no longer "how do I climb?" but "why does the model not understand me?" — a technical and informational question before it is a marketing one.

The first cause to check: your server may be shutting the door

Before the assistant reads a word of your site, it has to reach it. Some shared hosts enable off-the-shelf "bad bot" lists, and those lists file AI crawlers alongside the scrapers that steal content. When one is switched on, your server answers 403 to GPTBot, ClaudeBot and PerplexityBot, and you do not know.

This happened to our own site. The server answered 403 to eight crawlers — including all three of OpenAI’s — on every path, robots.txt itself included. In other words, the assistant could not even read the file meant to grant it access. Our .htaccess was entirely clean: the block came from a security tool at server level that we did not control, and it was lifted only after a ticket backed by a measurement taken from inside the server itself.

The lesson: on shared hosting, assume this is happening to you until you prove otherwise. It is the first check, before any work on content, because everything below it stays without effect while the door is shut.

Ask your developer to compare the server’s response to GPTBot against its response to a browser.
If the two differ, the problem is at the host, not in your site.
Look in your panel for Imunify360, ModSecurity or Bot Protection.
If the settings are locked at server level, the fix is a ticket backed by a measurement.

What the assistant does not read on your site

The assistant reads text. Anything that is not structured text passes it by and never enters the answer. These are the places your most important information most often disappears:

Note that this is neither a design nor a speed problem. Your site can be beautiful, fast and responsive, and your prices still not exist for the model because they sit inside an elegant image.

Prices inside images or graphics — neither selectable nor readable.
Service lists inside PDFs with no clear link pointing at them.
Information that appears only after JavaScript runs or a button is pressed.
Terms and service areas spoken in a video with no accompanying text.
An "about" page describing an ambition without saying what you sell, to whom, and where.

Contradiction is worse than absence

Faced with missing information, the model moves on. Faced with two contradictory facts from the same source, it loses confidence in the whole source. One price on your site and another on your Facebook page, or hours on Google Maps that disagree with your own: either pushes the model towards a clearer source, usually your competitor.

Four facts in particular must be identical, word for word, everywhere they appear: legal and trading name, address, telephone, opening hours. That is what the trade calls NAP, and it is the basis of entity recognition in local search and in answer engines alike.

Prices too. If you do not want to publish a fixed figure, publish a starting price with its condition stated plainly — "from this much for a landing page" — which beats a silence the model will fill with an estimate from somewhere else.

Who are you? The entity before the content

Before deciding to cite you, the assistant needs to know exactly who you are and to tell you apart from anyone with a similar name. That is what structured data does: a schema.org block in your page stating your legal name, your address, your languages, your service area, your services and their prices.

The strongest thing you can add is what is verifiable outside your own site: a commercial register number, a tax identifier, and official profiles linked to one another. A company registered under a number somebody can look up is not the same as a trading name on a page.

Then comes the identity of whoever writes. Content with no identifiable author weighs less than content signed by a person whose experience is stated and whose profiles are reachable. This is not a formality: it is one of the clearest differences between a source that gets cited and a source that gets read and forgotten.

llms.txt: a summary for assistants, not a guarantee

llms.txt is a proposal made by Jeremy Howard in September 2024: a single text file at the root of your site that presents your organization in plain language — who you are, where, what you offer, at what prices, and which of your pages are the reference ones. It is not a standard, and nothing guarantees an assistant will read it; its value is that it makes you write your facts once, in one place, where any tool that does read it will find them.

The most useful part is the section almost nobody writes: what you do not do. A model with no negative facts fills the gap by guessing, and credits you with a service you do not offer. One explicit line, in the file and on your pages, leaves less room for that guess.

But be clear: the file is a declaration, not a repair. If your server blocks the crawlers it will never be read, and if your prices contradict each other it will fix nothing. Its place is after access and consistency, not before them.

Measure instead of guessing

Do not start by fixing. Start by recording what the assistants say about you today, because that alone is the baseline that will prove something changed. Gather twenty to fifty questions as a real buyer asks them in your sector, not as you describe your service.

Put them to at least three assistants and note, for each: were you named? was the information right? who was named instead of you? Take a screenshot of every answer. That file is worth more than any general report on "improving visibility".

Then put the same questions again a month after the fix, and compare. If nothing moved, the problem runs deeper than the technical layer and needs content and outside mentions. If something moved, you have a number to build on rather than an impression.

"What is the best [your service] in [your wilaya]?"
"How much does [your service] cost in Algeria?"
"Compare [your company] and [your competitor]."
"Is [your company] a good fit for a twenty-person business?"
"What are the risks of working with [your company]?"

The first thirty days

The aim of the first month is not to dominate the answers, but to remove everything stopping you from being understood and to own a number you can measure against. Four weeks, each with one clear deliverable.

Week 1: ask the questions, archive the answers, and check crawler access to your server.
Week 2: lift any block, and build the entity data with your registration numbers, services and prices.
Week 3: move prices and services out of images and files into text, and unify your information across every profile.
Week 4: write FAQs in the buyer’s words, publish llms.txt, then measure again and compare.

Common questions

Is this a replacement for SEO?

No, and it does not work without it. The same foundation — a fast, clear, structured site — serves both. The difference is that SEO targets a position in a list, while this targets being understood by a model that writes the answer. Anyone who neglected the first will not succeed at the second.

Is adding an llms.txt file enough?

No. It is a proposal, not a standard, and nothing guarantees an assistant will read it. Where it is read, it declares your identity; it does not repair what stops you being understood. If your server blocks the crawlers it will not be read, and if your prices contradict each other between site and page it will not resolve the contradiction. The order is access, then consistency, then declaration.

How do I know whether my host blocks the crawlers?

Ask your developer to compare the server’s response to a request carrying the GPTBot signature against its response to a browser. If the first is 403 and the second 200, the block is confirmed. On shared hosting with no security tool visible in your panel, the setting is server-side and needs a ticket.

Do you guarantee ChatGPT will cite me?

No, and nobody can. The decision happens inside a model no provider owns. What can be guaranteed is that the obstacles come out — blocks, missing data, contradictions — and that a monthly measurement says what changed. Anyone selling you a guarantee of citation is selling what they do not own.

How long before it shows?

The technical fix takes days. When the assistants start mentioning you, nobody can say or promise: it depends on when their crawlers come back and on systems no provider controls. So record their answers before the fix, a month after, then every month. Anyone promising a date is promising what they do not control.

Does this matter in Algeria yet, or is it early?

Honestly: the number of people asking an assistant before Google is still smaller than the number searching on Google, and we do not expect that to flip this year. But the direction is clear and the share is rising, and the cost today is low because most competitors have not looked yet. Doing it now costs less than doing it in two years, when the ground is crowded.

Move from reading to a concrete step

Visibility in AI answer enginesA report of what the assistants say about you, the fix for what stops them understanding you, then a monthly measurement.The free diagnosticFifteen questions that surface your first gap, with no sign-up and no email.Website and online store developmentA site you own and administer, built to be read by people and machines alike.Digital Algeria 2030 guideThe reference page on the five axes and what they mean for your organization.