AI Search Optimization for SMBs: getting cited, not just ranked
On 22 July 2026, Google switched on generated answers in France. Press publishers had been warned by letter three weeks earlier. Small businesses found out in their analytics. Since then, some of your prospects read an answer before reaching your site, and leave satisfied. Ranking first is no longer enough. You have to be the source that answer cites.
generated answers went live in France
of tracked searches, four days later
outbound clicks without the summary on top
before a first measurable result
What changed on 22 July, and what it costs you
The shift was fast and it is still running. No setting to change, no choice left to anyone.
What it looks like on screen is simple. Someone searches for a provider, a solution, a price range. They first see a paragraph that answers, with three or four sources cited beside it. If satisfied, they do not scroll.
Four days after launch, a quarter of the French searches tracked already carried a generated answer. Semrush and Ahrefs measure different rates, which gives the measure of the uncertainty as much as of the phenomenon. Sources: Semrush Sensor and Ahrefs, 26 July 2026.
The measured consequence
Two neutral studies establish it. The Pew Research Center, a non-profit with no stake in search marketing, measured that a click on a classic result falls to 8 % of visits when a generated answer is present, against 15 % without.
A randomised, pre-registered experiment by Agarwal (Indian School of Business) and Sen (Carnegie Mellon) measured a drop of about 38 % in outbound organic clicks. Removing the summary when it sits at the top produces 88 % more outbound clicks, with no measurable effect when it appears lower down.
What it shifts for you
Your search work has not become useless, it has become insufficient. You can gain rankings and lose traffic in the same move, which had never happened before.
We anticipated this shift and published a guide to AI Overviews in France before they arrived. This page does not repeat it: it covers what we concretely do.
What GEO agencies sell, and why it does not hold
A new offer has appeared everywhere under the name GEO. Let us save you paying for air.
Magic markup does not exist
Google was explicit: search fundamentals remain applicable to generative features, and no specific markup or parallel method guarantees a presence in those answers.
Google went further and stated that chasing artificial mentions is ineffective too. Two official denials, on the two techniques most often sold.
The figures come from the sellers
You will read that assistant traffic converts nine times better than organic. It may be true. But the sources are agencies selling the service, on samples they do not publish.
We do not repeat them. We would rather show you what we measure on your own market, which has the advantage of being verifiable.
Billing that does not hold up
Published price lists from French agencies put the service at roughly 1,000 to 3,000 euros a month for an SMB, depending on whether it is billed separately or folded into search work.
We do not bill it separately, and that is not a discount. On Google, being cited follows directly from search work. Billing both apart would mean selling you the same thing twice.
What actually gets a company cited
Here is what we work on, and why.
A study by researchers from Princeton and IIT Delhi, published at the ACM SIGKDD 2024 conference, compared nine rewrites of the same text to see which ones got picked up most in generated answers. Three levers stand out: citing credible sources, backing the point with expert quotations, and giving figures. Keyword stuffing, on the other hand, pushes visibility down. The measured gain goes up to 40 % on the researchers test bench.
Figures that are true
One point the study does not test, and that we add: figures are only worth something if they are true and attributed. A made-up number that gets picked up is a reputational liability, not a win.
Structure and expertise, our own method
Two further levers belong to our own practice rather than to that paper: a clear structure, and demonstrated expertise. They are not measured by the study, we present them as what we do. What reads well for a machine reads well for a busy executive.
Questions asked, not keywords placed
People no longer type expressions, they ask whole questions with their context. We write your pages by asking the question as it is asked, with the answer directly beneath.
A presence off your own site
For a language model, a mention of your company in a comparison or a professional discussion is already usable information, even without a clickable link.
One necessary caveat, and it follows from what Google said: this means natural mentions, not manufactured ones. Chasing fabricated mentions is precisely what the engine declares ineffective.
Which engine to target first as a French SMB
It is the first question business owners ask, and the answer varies by market.
Google, first and without hesitation
By far the largest volume in France, and the simplest to work, since citation follows from your organic ranking. The work to do is search optimization. Anyone selling you a special device for it is selling you a duplicate.
ChatGPT, second and for other reasons
Growing fast in France, and running on a different index. Content can be cited there without ranking in classic results. Worked less through position, more through mentions: slower, but more durable.
Perplexity and Gemini, by sector
A minority on the French market, but not negligible everywhere. We do not target them as a priority, we check whether they matter for you before deciding. That takes a few days and avoids effort that would return nothing.
The overlap trap
Overlap between the sources these engines cite is low. There is not one generative visibility to win, but several, and the same work does not win them all.
What you can check yourself, for free
Before paying anyone, including us, you can find out where you stand. Three checks, one hour, no tools.
Are your queries affected?
Type five searches your prospects would make. Not your name: the questions they ask before they know you. Count how many show a generated answer at the top. Three or four out of five, and your visibility is already substantially exposed.
What does an assistant say?
Open ChatGPT and ask the question a customer would ask. See who gets named. If it is your competitors, you already know what is missing. Then ask about your own name: what the model says about you is sometimes wrong.
Do your best pages lead somewhere?
Open your three most-visited pages and look for the link to your offer. Many well-ranked pages describe the service and offer no clickable link. The visitor reads, understands, and leaves. It takes minutes to fix.
How we work on this at Synerium
Four things, and they owe as much to our model as to our craft.
We measure your exposure, monthly
Which queries in your market show a generated answer, which do not, and how it evolves. That data exists and is accessible. It tells you which share of your visibility is exposed, instead of assuming it.
We track your presence in assistants
Querying conversational engines by question phrasing rather than keywords, because that is how people address them. You then know whether your brand appears, on which questions, and alongside whom.
We arbitrate what is still worth it
Some queries no longer return much even in good position. Others remain intact because they involve a choice, a quote, a human contact. Separating the two makes an immediate difference to your numbers.
We gain nothing from selling you a rebuild
Our clients pay a monthly subscription, not a project. When Google changes how it answers, we adapt your pages the following month, with no quote and no negotiation. That is the whole point of Synerium Infinity.
The mistakes we see most often
Paying for a technical device
The costliest error, because it feels like action. Google denied that any of this guarantees a presence. The money is spent, visibility has not moved, and nobody notices for months.
Targeting every engine at once
Splitting effort across four grounds without knowing which matter dilutes an already limited budget across three that will return nothing.
Publishing more, assuming volume compensates
A generated answer draws on a handful of sources, not hundreds. One more piece repeating what twenty others already say has no reason to be selected.
Neglecting what is said about you elsewhere
On conversational engines a natural mention in a comparison weighs more than a perfectly optimised page at home.
Acting without measuring
Many companies discover after six months that they worked ground that did not concern them. The three free checks above take an hour.
What it looks like in your subscription
Search work, editorial production, structuring for extraction, measurement and technical fixes are part of the same monthly subscription, from the Premium plan at 2,990 euros excl. VAT per month on a three-month commitment, or 2,492 euros on a twelve-month one.
One service, or a whole department
The comparison that matters: the market bills 1,000 to 3,000 euros a month for the GEO service alone. At Premium, the same budget covers design, SEO, paid search, content, newsletters, social ads and landing pages.
No line, no quote, no surcharge
A typical month can see two pages restructured, one long-form article, the reading of your exposure to generated answers and a fix to your internal linking. Without any of it having to be quoted.
What is AI search optimization, or GEO?
Work aimed at getting your brand cited in answers generated by engines: the summaries Google has shown in France since 22 July 2026, but also ChatGPT, Perplexity and Gemini. The discipline was formalised by researchers from Princeton and IIT Delhi, in a study published at the ACM SIGKDD 2024 conference, which compared nine rewrites of the same text to see which ones got picked up most.
Does classic SEO become useless?
No, it becomes insufficient. On Google, being cited follows directly from your organic position: it's the same work. What changes is the trade-off, since second or third place now returns almost nothing.
What does it cost at Synerium?
Nothing beyond your subscription. Published price lists from French agencies put the service at roughly 1,000 to 3,000 euros a month for an SMB, depending on whether it's billed separately or folded into search work. We don't bill it separately, because on Google it's the same work with a different trade-off.
Do I need particular technical markup?
No, and Google explicitly denied it. Google has also stated that chasing artificial mentions is ineffective. Be wary of offers selling a file or a declaration as the solution. What counts is editorial structure, sourced data and your genuine presence off your own site.
Which engine should I target first?
Google first, without hesitation: by far the largest volume in France, and the work there is classic search optimization. ChatGPT second, growing fast and running on a different index. Perplexity and Gemini remain a minority, but their weight varies by sector, so we check whether they matter for you first.
How long before a result?
Three to six months minimum, longer if your domain is recent. Any promise of results in weeks should alert you.
Can you guarantee ChatGPT will cite me?
No, and nobody can. Models do not publish their selection criteria. What we guarantee is doing the work that improves the odds, measuring it monthly, and showing you the results unarranged.
What's the difference between Google summaries and ChatGPT?
Google's summaries rest on its index and your organic positions. ChatGPT, Perplexity and Gemini run on different indexes, where content can be cited without ranking in Google's top results. Overlap is low.
Do brand mentions count without a link?
On conversational engines, yes, and that is a major difference from classic search. A mention in a comparison, a press article or a professional discussion is already information for a language model. One important caveat: this means natural mentions. Google has stated that chasing fabricated ones is ineffective, and we do not manufacture them.
How do I know if my queries are affected?
You can check it yourself in ten minutes: type five searches your prospects would make and count how many show a generated answer. For full tracking, we measure it monthly with the evolution over time.
Want to know where you stand?
The full reading of your exposure to generated answers is included in your first month. It will tell us together whether the subject is urgent for you, or whether it can wait.