1 · Google let the habit settle in, then answered everywhere at once
Late 2022, ChatGPT comes out, and a new habit takes hold: asking a machine that answers in sentences. Google watches, tests its Search Generative Experience in the lab, and waits. The response takes eighteen months to leave the lab: May 2024, AI Overviews (the AI summaries displayed above the results) reach the United States. On March 26, 2025, nine European countries receive them at once, Belgium included. France waits sixteen more months, the time to strike a deal with its press publishers: rollout on July 22, 2026.
Market | AI answers since | Lived with, as of August 4, 2026 |
|---|---|---|
United States | May 2024 | 27 months |
Belgium, with 8 EU countries | March 26, 2025 | 16 months |
France | July 22, 2026 | 13 days |
Now that the machine is running, it moves fast, and in bursts. Semrush tracks the share of queries showing an AI summary across 10 million American keywords: 6.5% in January 2025, 24.6% at the July peak, then a pullback around 15% by year end. Google adjusts the dial, query by query, and nobody knows where it will settle. What we do know: the movement does not reverse.
2 · The time delta: accumulated traces weigh more than your latest tweak
A model answers with what it knows: its training corpus, plus whatever its index feeds it. An established brand drags years of articles, reviews, press mentions and archived pages behind it. Those traces corroborate each other, and corroboration is what makes the citation. The longer a brand has existed and the more traces it has left, the stronger the effect. An Ahrefs study of 75,000 brands measures the same mechanics: accumulated brand mentions rank among the signals most correlated with a brand's presence in answers. We see it market after market; we cannot quantify it cleanly yet, so we state it as an observation, not a law.
We live this delay from the inside. Our own site is a few months old, with an authority scored at zero by the tools at launch, Google impressions climbing, AI citations still rare. Nothing abnormal: seniority cannot be decreed. It works like a sourdough starter, it asks for time, not talent.
What this means for a young domain: time cannot be caught up, it gets compensated. Through data nobody else owns, through mentions on surfaces that are already old and credible, through patience. The leads below follow from that.
3 · Ranking first on Google does not mean being cited by AI
Google rankings and citations in AI answers are two different games. We see it on our own pages: decent positions on targeted queries, and rare citations in assistant answers. The overlap between the two is small, the full numbers sit in our article GEO is just SEO.
The reason is mechanical. An assistant draws from two reservoirs. Its internal library first, the corpus the model was trained on: it is not built live, it freezes at training time and updates in cycles, months apart. Its search index second, consulted the moment you ask. Your page published yesterday can be indexed tomorrow and stay unknown to the library for months: entering a model's memory is often slower than classic SEO. Two speeds, two work sites.
The practical consequence: your ranking report says nothing about your presence in answers. Steering by rankings alone is watching the rearview mirror while the road splits in two. You need eyes elsewhere, and that is what the next leads are about.
4 · Our server logs: the llms.txt file gets almost no visits
The llms.txt file, a plain text file at the root of a site telling AI engines what you would like them to pick up, is presented almost everywhere as the foundation of AI search optimization. We have published one for months. Our server logs, the visit journal every server keeps, are stubborn: the file is almost never fetched. Not by OpenAI's robots, not by Anthropic's or Perplexity's. The finding reaches beyond our servers: at Google, John Mueller compares llms.txt to the keywords meta tag, which no engine has read in twenty years, and Google's official documentation on AI features states that no extra machine-readable file is needed to appear there. We keep ours, thirty minutes of work, zero risk. We just refuse to sell it to you as a pillar.
The lesson goes beyond this file. A new terrain attracts miracle-ingredient sellers: the tag that does everything, the secret setting. When someone promises you guaranteed citations through a technical trick, ask for their logs.
Observing the answers themselves stays just as uncomfortable. They vary day to day, account to account, and dedicated tools are barely out, Ahrefs' Brand Radar and Semrush's tracking first among them. We cross this young tooling with manual tests, owning the holes in the net.
5 · Check the indexes AI engines pull from, not just Google
Every assistant leans on a search index, the copy of the web a search engine maintains. ChatGPT queries Bing's, integrated by Microsoft since 2023. Claude leans on Brave Search, documented by TechCrunch, with Anthropic listing Brave among its subprocessors. Google's AI answers draw from Google's index. A page absent from these indexes will be cited by no answer, whatever its quality.
The lead takes one hour of work:
- Google Search Console: the base, impressions and index coverage.
- Bing Webmaster Tools: free, it imports your settings from Search Console in a few clicks. It is the index feeding ChatGPT, and almost nobody looks at it.
- Brave Search: no console to date. Type `site:yourdomain.be` plus your key commercial queries, and check you exist there.
Also check the reverse direction: that your door is not closed to AI robots without you deciding it. Cloudflare blocks them by default for its new customers since July 2025, and an inherited firewall rule can do the same silently. The zero-cost complement: IndexNow, the protocol Bing listens to, notifies the index of every new page in one request; we send it at every publication. You know where you exist, where you are missing, and who you let in.
6 · External consensus: mentions, partnerships, directories, YouTube
Before citing a name, an AI answer corroborates independent sources. That corroboration is built off your site: a solid domain authority (the credibility score SEO tools compute from the links pointing at you), consistent brand mentions, same name, same activity, same address everywhere, real partnerships that leave public traces, established directories. Our first lived-in link is of that kind: a listing on Digital Wallonia, the regional digital directory, a domain far older and better credited than ours.
Platforms weigh heavily in that corroboration. The video blocks in answers almost always feature YouTube, Semrush observes it across its 10 million keywords. Reddit sells access to its content to the models, a deal with Google reported at around $60 million a year. When an engine pays to draw from a surface, the conclusion draws itself: being present there counts. Wikipedia closes the list, precious and codified at once. Do not rig a page to your glory there, it will be removed; a brand settles in through its sources.
A restaurant's reputation is made in the guides and the dining room, never in its own kitchen.
7 · Reallocate the editorial effort toward what AI must cite
The writing effort stays, its target moves. Generic informational content, the definitions and generalities found on ten sites, is the first thing summarised without a click: in the United States, when an AI summary shows, 8 visits out of 100 still click a result, versus 15 without a summary. Pew Research measured that behaviour across 68,879 real searches. Writing yet another "what is X" page means feeding someone else's answer.
Where to move the effort: toward pages that precede a decision, honest comparisons, prices, concrete cases, and toward the data only you own. Our bet this year is of that type: an X-ray of 579 e-commerce shops in the Liège area, measured shop by shop. That data exists nowhere else; an AI that wants to answer seriously on the topic has to go through it. The way of writing to get picked up, stated definitions, clean entities, information in the visible text, fills a page of its own: our AI search optimization (GEO) method.
8 · Measure without fooling yourself
Here is what we watch every month, without claiming the full picture:
- referral traffic from assistants in GA4, those visits arriving from chatgpt.com or perplexity.ai, rare but real, and often qualified;
- impressions and queries in the consoles from lead 5, Google and Bing side by side;
- manual tests on key commercial questions, noted, dated, redone at regular intervals.
The consoles are catching up, by the way. Search Console launched in early June its performance report for generative results: AI Overviews and AI Mode impressions finally separated from the rest, but no clicks, no queries, and a still gradual rollout. Bing Webmaster Tools opened its AI Performance report in February, Copilot citations included. On a brand our size, these reports still show very little: we watch them without expecting the light just yet.
None of this properly measures a "share of voice in AI". Tools promising it extrapolate, and honesty about measurement is part of the job: the same reflex has guided our tracking & measurement work from day one.
Sixteen months of head start mostly taught us to separate what can be observed from what gets sold. If your business lives on local, transactional queries, the answer block barely touches you yet: start with the classic foundations. If your brand is young, compensate time with data and mentions, the rest asks for patience. The perfect bake does not exist yet, but the batches keep improving: we will keep weighing, tasting, adjusting. As always with SEO.
