Being cited by an AI: what actually changes

Two stories circulate on this subject and both are wrong. The first says traditional search is dead and everything must be redone. The second says these are toys and there is nothing to do.

The published figures settle it, and they give a more interesting answer than either: traffic from assistants is tiny and it is growing very fast. Both are true at once, and it is the combination that dictates what to do.

This article explains what those figures actually measure, why the sources publishing them disagree with each other, and what can be done today — with no budget and no rebuild — to be the page an assistant cites rather than one of the ten it ignores.

There is no Algerian data on the subject. The figures below are global, and each chart says so.

Here is the point almost no article on the subject mentions, and it is the most useful one. The two main public sources give different numbers for the same thing: Similarweb puts ChatGPT’s share at 74.78% and Perplexity at 7.23%; SE Ranking gives 79.8% and 11.8%.

These are not errors. They are different methodologies — different panels, different definitions of what counts as a referral visit, different samples of sites. Neither is wrong and neither is the truth.

So take the order of magnitude and never the decimal. An article quoting you "79.8%" without saying where it came from, and without mentioning that another serious source says 74.78%, is giving you a precision that does not exist.

A practical consequence for reading any article on this subject, including this one: always ask what period it covers. These shares move several points in twelve months, which means a figure published two years ago describes a market that no longer exists. The date here is part of the data, not a footnote.

A consequence for reading this very article, which we accept: the figures above will be out of date before most readers reach here. That is why each chart carries its year, and why this section’s conclusion is about a direction rather than a level.

A question as the section heading, and the answer in the two sentences that follow it. No introduction, no scene-setting, no "before answering, it is important to understand that". The answer first, the nuance afterwards.

Checkable facts rather than adjectives. "We are the sector leader" is extractable by nobody and nothing; "the online payment acceptance file runs in weeks rather than days" is, because it is a claim that can be true or false.

And figures with their source and their date, as in this article. That is what separates a citable passage from a plausible one, and it is exactly the standard a competent human reader applies.

A simple way to check that a passage is extractable: read it aloud imagining it is the whole answer. If you need to have read the previous sentence for it to make sense, it is not standalone. That test takes ten seconds per section and is more reliable than any tool sold for the purpose.

There is no reliable, complete tool for this yet, and be wary of anyone claiming otherwise. What exists is more manual and is enough: ask several assistants the questions your customers actually ask, and look at what is cited.

Do it in your customers’ words, not yours. And redo it monthly: answers change, including for the same question asked twice.

Look at your traffic statistics too. Traffic from these platforms shows up there like any other referral source, at small volume; if it is zero while your competitors are being cited, you have your answer.

Finally, write your reference answers down somewhere, with the date. Without that history you will not be able to say whether your position is improving: answers change month to month for reasons that have nothing to do with you, and memory will mostly retain the time you were cited.

Look too at whether competitors are cited on questions you handle better than they do. That is the most actionable information this record produces: it does not tell you to write more, it tells you an existing page is not phrased so it can be picked up, which is a problem of form and takes an hour to fix.

We start by establishing your baseline, which exists in no tool: we ask several assistants the questions your customers actually ask, in their words, and note who gets cited. You receive that record, dated. It is the only document that will make it possible, six months from now, to say whether anything moved.

We then rework your most-visited pages so each section answers a question in its first two sentences, so claims are checkable rather than adjectival, and so every figure carries a source and a date. That is the work described in this article’s tenth section, and it happens on your existing pages rather than on new ones.

We do not bill it as a service separate from search work, because it is not one: these are the same pages, improved for the same reasons, and classic traffic remains what brings you the rest. An agency selling you a distinct "AI optimisation" line is selling you the same work twice.

And we do not promise you will be cited. Citations cannot be bought, there is no auction, and no tool today tracks this completely — the FAQ below says so too. What we commit to is the monthly record and the work on the pages; the rest depends on what you wrote.