Digital marketing could tie every euro to a click. With AI, we are heading back to impression marketing.

1 · What Nadège could prove

When Nadège started out, in 2014, traditional marketing spoke in reach: a radio spot was sold on audience, a billboard on estimated passers-by. You knew how many people might have seen the message, rarely what they did with it.

Digital gave her the opposite: tie every euro to a query, every query to a click, every click to a sale. The Google Ads search terms report showed her what people had typed before clicking, for instance "vélo électrique cargo liège". Google Analytics followed the visit all the way to the test-ride booking, and the CRM said who had bought. She cut the campaign that brought nothing in to double the one that worked, with the figures to prove it to her managing director.

AI is breaking that chain link by link: the query gets lost in a conversation, the click becomes optional, and measurement tools have not kept up.

2 · One sale, claimed four times

Take Julien, 46, a typical customer: he lives in Seraing and is looking for a bike to ride up to work without arriving drenched.

1. On Instagram, he sees a Liège videographer climb the Côte de la Redoute on one of the shop's bikes, in a collaboration paid for by Nadège.
2. He asks ChatGPT which e-bike to choose for fifteen kilometres of hills a day. The answer names three brands, two of them sold in the shop; he does not click any link.
3. He types "vélo électrique liège" into Google. An AI Overview sums up the criteria to look at, citing a competitor.
4. He searches for the shop's name, clicks the Google Ads brand ad and signs up for the newsletter.
5. Two weeks later, the video crosses his feed again, as an ad this time. The next day, the email about Saturday test rides gets him to book a slot.
6. On Saturday, he tries three models and buys a €3,400 bike in the shop.

Each tool sees its own piece and counts the booking as its own.

Add it up and you find the gap Nadège was looking at on Monday: 88 sales claimed between them by Google Ads, Meta and the email tool, against 52 real ones. The decisive step, the one where Julien settled on two brands, leaves no trace at all.

3 · Google builds in AI, its tools lag behind

In Search Console, on "vélo électrique liège", Nadège's impressions go up and her clicks go down: the AI Overview answers before the first link. AI Overviews has shown in Belgium since March 2025 (our article on AI Overviews in Belgium took stock after sixteen months), and Google's tools still cannot say what these answers bring in.

The Search Console section devoted to generative AI gives impressions for pages cited in AI Overviews and AI Mode, with no clicks or queries, and a single position for every link in the same answer. An impression counts as soon as the answer appears, even if nobody scrolled down to your link. In September, John Mueller, Google's spokesperson, admitted that this position "is hard to do in a way that makes it useful". GA4 does no better: it counts a click from AI Overviews or AI Mode as ordinary organic search.

The results themselves move without explanation: Google updates its algorithms "continually", including through smaller core updates it does not announce. Between 21 September and 12 October, our site went from 93 to 71 indexed pages out of 123, while our /squad page climbed to third place on "agence marketing digital liège". Since we had changed the site in the meantime, there is no telling our changes apart from Google's.

4 · ChatGPT publishes nothing, tracking tools guess

Julien asked ChatGPT his question, and Nadège will never know. ChatGPT, Claude, Gemini and Perplexity offer no equivalent of Search Console: no way of knowing which questions a brand appears on, or next to which competitors. Only Bing counts a site's citations in Copilot, without views or clicks.

GA4 did add an AI Assistant channel in May, which recognises visits from ChatGPT or Gemini by their referrer or by their tracking parameter. ChatGPT's Atlas browser often strips the former, and in the week of 8 October, the latter vanished as well.

Nadège first thought ChatGPT had stopped citing her: her visits from the assistant were collapsing while her direct traffic climbed. Since the release of GPT-6, ChatGPT no longer put the `utm_source=chatgpt.com` that signs a visit's origin into its links, nor did it pass on a referrer: GA4 therefore filed those visits as direct. Depending on the measure, traffic attributed to ChatGPT fell by 22% to 60%; Cedric Huesler, at Adobe, counted across some 500 pages 31,000 fewer ChatGPT visits a day and 26,000 more direct ones. OpenAI confirmed a bug to him on 14 October, without any public announcement, and the parameter came back the next day. No report had raised an alert.

To fill the gap, Nadège tried an AI tracking tool, of the Semrush or Peec AI kind, which regularly puts a list of questions to the assistants and counts how often the brand appears. Except a question is nothing like a keyword: it can be phrased a thousand ways, and two identical runs do not always give the same answer. Her panel has thirty questions in French, while some of her customers write from Eupen in German, or from Limburg in Dutch. Our BrandMonitor report in September gave 91 out of 100 when the question named HeySquad, 2 out of 100 without the name: two accurate figures that do not measure the same thing. These tools sketch a trend if you keep the same panel, never a performance.

5 · ChatGPT Ads: AI advertising goes back to impressions

At the end of August, Nadège was offered a ChatGPT Ads test, launched in Belgium with a self-serve campaign manager on 31 August. The audience is huge, 1.2 billion weekly users according to OpenAI, but the buying model takes us backwards: during the pilot phase, these ads were sold per thousand impressions, like a radio spot, 60 dollars at launch and down to 25 dollars in April according to Digiday.

Nadège would not see the questions that trigger her ad, only totals. Conversation privacy justifies it, but for her, it amounts to buying Google Ads without the search terms report.

The reported figures do not always add up. PPC ONE compiled 49 published cases and 47 advertiser reports: one campaign manager shows 36 conversions for two real enquiries that came from somewhere other than ChatGPT, another 57 clicks for fewer than 20 visits in GA4. Among the four advertisers comparing both ad platforms on their own account, ChatGPT converts far less than Google, for instance 0.46% against 5.3%.

OpenAI admits that its work on incrementality, meaning the share of sales the ads actually caused, "is only just beginning". The company is exploring geographic experiments: switch the ads on in one area, not in the other, and compare. It is the method of traditional marketing, applied to the newest ad platform on the market.

6 · A dozen channels, one Nadège

While measurement recedes, Nadège's workload piles up. In 2014, she ran Google Ads, a bit of Facebook, the website and a newsletter. Since then, Instagram and TikTok have joined, LinkedIn for the bike leasing offered to companies, SEO, visibility in AI answers, ChatGPT Ads and three collaborations with creators per season, each with its own interface, metrics and formats. On Monday, Nadège compiles six exports into a spreadsheet; the rest of the week, she decides where to invest, adjusts the campaigns, briefs the creators and adapts every visual into a dozen formats. Before the budget even comes into it, the mental load is what holds her back.

Some platforms promise to track everything in a single dashboard. Yet each one sees through its own integrations, and several belong to a company that also sells advertising space.

At HeySquad, we do the opposite: each tool stays in its place, and AI makes the link. MCP, short for Model Context Protocol, lets an assistant such as Claude query GA4, Google Ads, Meta Ads, Search Console or the CRM directly. Nadège would ask her Monday question in plain words, "what brought in September's test-ride bookings?", and the assistant would put the figures from each source side by side. It does not solve attribution, it removes the hours of compiling. At Ecostal Group, the same AI-equipped principle, applied to production, brought the build of a complete campaign down from 14 hours to 2.

Before adding a channel, the question is what role it plays, and how to judge whether it fills that role.

7 · What Nadège must do in 2026 to take back control

Nadège will not get the 2014 chain back. She can build another one, closer to traditional marketing, that starts from the sale instead of the click.

We no longer link every click. We link every sale to what came before it.

Six moves are enough to get started. Most require no new tool, only regularity.

1. Ask where the contact came from. One field in every booking form, and the same question asked in the shop at the test ride. Julien would have answered "a video on Instagram, then ChatGPT".
2. Make the CRM the referee. Record every sale with its declared source, then feed it back to Google Ads, Meta and ChatGPT Ads so their algorithms learn from real sales.
3. Own your measurement. Server-side tracking routes the data through your own server rather than through the visitor's browser alone. At Ecostal Group, six brands across three countries, it recovered 22% of conversion value the native tools could not see. Details in our article on first-party measurement.
4. Read direct traffic and brand searches together. If both rise with no campaign to explain it, it is often a sign that AI is citing you. If direct rises alone, look for a tagging problem first, like ChatGPT's in October. GA4 annotations are enough to date every site change, every campaign and every Google update.
5. Control what AI reads about your brand. Assistants build their answers from what they read elsewhere: directories, profiles, reviews, articles. In our BrandMonitor report, the AI's reservations about HeySquad came from incomplete or outdated third-party listings, and from details that varied from one page to the next. For Nadège, that means an up-to-date Google Business profile for each shop, recent reviews, the same opening hours and addresses everywhere, and pages that answer the questions Julien actually asks. It is the foundation of AI search optimisation.
6. Cap the presence budget, and test it. Decide in advance the share of budget accepted without a direct link to sales, then switch it on and off by period. With two shops, Nadège can even run her own geographic experiment: ads around one shop, not around the other, then compare sales.

Our tracking and measurement page details this foundation, with GA4 and dashboards that cross-check the sources.

8 · Invest, yes. But how much?

Nadège's managing director has rolled over the 2027 budget unchanged. In reality, media costs rose by 4.3% in 2025, after about 5% in 2024, that is 1.7 times general inflation, as EPIC points out in its analysis of the 2027 marketing budget (in French). A budget that is simply maintained buys 4 to 5% less visibility every year.

Drowning people in impressions is therefore not a good idea. Yet it is the natural slope: an ad platform sold per thousand impressions earns more when you buy more, and blurry measurement benefits whoever sells the space first. Anyone who does not steer properly no longer knows when, how or where to invest. By default, they buy volume.

Cutting these channels would be no better: Julien settled on his two brands in a ChatGPT answer, long before walking into the shop. The real question is the amount, and three benchmarks hold up. The presence share is decided in advance, and the weaker your measurement, the smaller it stays. It is indexed on your own cost per thousand impressions, because a frozen budget shrinks without anyone deciding it. Finally, it is judged over time: according to Thinkbox and Ebiquity, quoted by EPIC, one euro invested in media returns €1.87 in profit within three months and €4.11 once long-term effects are counted. Steering on cost per acquisition alone means missing a good part of what presence brings in.

These are early analyses, on ground that shifts every week: the tools will probably end up tracking better, because advertisers are demanding it, and nobody knows when.

Next Monday, Nadège will not walk in with four figures that do not add up. She will bring the CRM sales, the source declared by customers, the brand search curve, and the share of budget devoted to presence, tested from one shop to the other. Her managing director will not get the 2014 answer, he will get what he needs to decide.

An honest caveat: none of this applies yet to a small shop starting out with a handful of sales a month. At that stage, a question about where the contact came from and a well-kept CRM are enough. For an SME that invests in acquisition every month, it is the only way to know what that money brings in. Book a meeting to talk it through.