In short
- What works today is narrow and operational. Producing content at volume, personalising messages, automated ad bidding, analysing customer verbatims, first-line support.
- What does not yet deliver is autonomy. No small company should let a tool decide a strategy, a media budget or a published piece of content on its own.
- The most structural change is happening in search. A growing share of questions gets an answer synthesised by an assistant, with no site visit: your challenge becomes being a citable source, not just a well-ranked page.
- The input data decides the quality of the decisions. A model plugged into faulty measurement produces faulty calls, simply faster and at greater scale.
The question we get in meetings has changed. In 2023 people asked whether AI would replace marketing teams. Since 2025 they ask instead why three AI subscriptions have not moved revenue at all.
In most cases we see, the tools were plugged into isolated tasks: write faster, summarise a meeting, generate a visual. The gain is real, but it stays time saved by one person, not margin gained by the company.
Three years of hindsight now allow some sorting. Some uses produce a result you can read in an indicator, others are still sales talk, and one underlying shift passes unnoticed: the way your customers look for information is changing.
What AI genuinely changes, and can be measured
A use deserves your budget when you can answer two questions: which indicator moved, and over what period. Five uses pass that test in a small company.
Producing content at volume
This is the most mature use. Product pages, ad variants, translations, versions of the same message by segment: assisted generation turns a task from several days into a few hours, and that shows directly on an editorial schedule.
The condition is not negotiable: a human reads and approves before publication. The model knows neither your stock, nor your real lead times, nor your contractual commitments.
Personalising without multiplying workflows
Subject lines adapted to the segment, product recommendations based on purchase history, follow-ups whose tone changes with customer tenure: these mechanisms existed before, but each variant had to be built by hand.
The model produces the variants, your marketing automation tool distributes them. The result is measured on open rate, click rate and revenue per email sent, against the generic version.
Steering ad bidding
Performance Max at Google, Advantage+ at Meta: since smart Shopping campaigns migrated to Performance Max in 2022, bidding, placement and delivery are handed to the model. On sufficient conversion volume it beats manual steering, because it continuously processes signals no team could process by hand.
What these campaigns do not do: decide what to sell, at what price, with what promise, or check that the conversion reported to them corresponds to a real sale with positive margin. That is exactly where budgets get lost, and it is human work.
Reading your customers’ verbatims
Reviews, support emails, free-text survey answers, product return reasons: this material already sits in your inboxes and exports. A model classifies it by reason in minutes and surfaces the irritants by frequency.
It is often the fastest gain to obtain, because the data exists and nobody was using it for lack of time. A return reason that comes out top across three months of history is worth more than an ideation workshop.
Handling first-line support
Order tracking, availability, return conditions, delivery times: an assistant connected to your data answers these correctly, around the clock, in several languages. The load taken off the team is visible in the first month on ticket volume.
The point to watch is the handover. An assistant that cannot say it does not know, and has no escalation path to a human, damages your customer relationship instead of improving it.
What does not work yet in a small company
Three promises circulate a great deal in 2026. We have not seen any of them hold in a company under 250 people.
| The promise | What we observe | What works instead |
|---|---|---|
| A model builds your strategy | The plan looks coherent, but it ignores your cash position, your production constraints, your local competitor, and what you have already tried | The model produces the options, the decision stays human |
| An agent runs the ad budget alone | It executes an instruction, it does not judge that a rise in acquisition cost is acceptable because you are launching a range. The loss is counted in days | Automated bidding, with a check that the reported conversion is a real sale |
| Generated content published without review | The factual error goes unnoticed in the writing and becomes visible to the customer: a specification, a price, a warranty | Production at volume, read and approved before going live |
What these three failures have in common fits in one sentence. AI is excellent at producing options, mediocre at choosing between them when the choice depends on context it does not have. A serious digital transformation starts by mapping that context, not by buying licences.
Your visibility no longer plays out only on the click
Here is the point we consider most decisive for a small company, and the least addressed. A growing share of questions asked online no longer ends on a list of links, but on a written answer: a generated summary at the top of Google’s results, a conversational assistant’s reply, a synthesis produced by an engine built into a browser.
The consequence is mechanical. If the user gets their answer inside the interface, they do not visit your site, and your audience falls even when your ranking position stays identical. You can keep good search rankings and lose traffic.
The reversal is this: the goal is no longer only to rank well, it is to be the source the model reuses and cites. Those are two different exercises, with different criteria.
What being a citable source means
A model looks for short, self-contained, attributable passages.
| What the model looks for | What that requires on your page |
|---|---|
| An answer extractable as is | Three lines answering the question at the top of the page, before any development |
| A hierarchy before prose | Explicit headings, questions phrased the way users ask them, marked-up data |
| A real, identifiable entity | Name, address, authors, field of expertise, consistent mentions everywhere |
| A fact it can attribute | Dated, sourced, updated figures. With no date and no source, a figure is ignorable |
| Information confirmed elsewhere | The same data on your site, a directory, a business listing, a third-party publication |
This work largely overlaps with what we already do in search and content, with the cursor moved. Structure and factual precision take precedence over word count and keyword linking. It is the core of our artificial intelligence expertise.
The loop risk: when everyone produces the same content
If your sector massively adopts the same tools, it massively produces the same content, trained on the same sources. The result is convergence: the same advice, the same plans, the same phrasing, across dozens of sites.
In that context, differentiation no longer comes from the ability to produce. It comes back to what cannot be generated: your proprietary data, your real client cases with figures, your field experience and the mistakes you made.
It is also what makes a source citable by an assistant. A piece of data only you hold, a dated field report, a comparison you actually ran: a model has no reason to cite a page repeating what ten thousand other pages already say.
What to take away Differentiation no longer comes from the ability to produce: a whole sector produces the same content from the same sources. It comes from what cannot be generated, your proprietary data, your client cases with figures, your field mistakes. That is what gives an assistant a reason to cite you.
An AI plugged into wrong figures decides wrongly, faster
This is the least spectacular and most expensive subject. Automated campaigns learn from the conversions you report to them. If that reporting is incomplete, duplicated or biased, the model diligently optimises towards the wrong target.
The causes are known: badly handled consent, missing or duplicated tags, ad blockers, browser restrictions on third-party cookies, offline purchases never reconciled. The typical symptom is a persistent gap between the revenue shown by the ad platform and the one in your back office.
Before investing in an AI tool, measure that gap. We cover the mechanism and its remedies in our article on server-side tracking and measurement reliability. For companies wanting to test server-side collection without commitment, DataFirefly Server-Side, published by Datafirefly Limited which belongs to the same group as Dotsland, offers a free plan of 10,000 requests a month.
The mistake that costs most Plugging automation into faulty measurement. Campaigns learn from the conversions you report: if that reporting is incomplete, duplicated or biased, the model optimises towards the wrong target, faster and at greater scale.
Where to start this week
A useful approach fits in six steps, doable with no transformation budget and no hiring.
- Pick one single process, the one costing you the most time. Writing product pages, sorting incoming email, repetitive support replies: one process, not three.
- Measure its current cost. Hours per week, or cost per unit produced. Without that starting value, you will never know whether the tool returned anything.
- Check the input data before the tool. Compare a month of your ad platform’s conversions against your back office orders. A gap above 10% is fixed before anything else.
- Test over four to six weeks, with a named owner. One identified person, a written scope, a review date. A test with no owner dies on its own.
- Put the human safeguard in writing. Nothing published, sent or spent without approval from a named person. The rule must be a document, not an intention.
- Decide at the review: keep, adjust, stop. Compare the process cost before and after, licences included. If the gain is not visible in an indicator, the tool does not deserve to be kept.
This sequence gives a repeatable frame: one process, a starting measurement, a bounded test, a decision. It is the same logic we apply in our digital marketing strategy engagements, where the constraint is never the number of ideas, but the ability to prove one paid off.
Frequently asked questions
Can AI replace my agency or my marketing lead?
No, not in 2026. It replaces execution tasks: first drafts, variant production, classifying customer feedback, first-line replies. It does not replace judgement, which requires knowing your margin, your cash position, your local competition and your production constraints.
How do I know whether an AI tool actually makes money?
Measure the process cost before deploying the tool, in hours or euros per unit produced, then measure again after four to six weeks. Compare the difference to the total cost of licences and setup time. If you cannot name the indicator that moved, the tool has not proved itself.
Is search marketing still worth it if people use AI assistants?
Yes, but with a broader objective. Assistants rely on web pages they must be able to read, understand and attribute. Content that is structured, factual, dated and consistent with your other online presences now serves two audiences: the classic search engine and the model composing an answer.
Is AI-generated content penalised by Google?
Google assesses the quality and usefulness of content, not how it was produced. What gets penalised is content produced at scale with no added value for the user. A text assisted by AI, reviewed, enriched with proprietary data and real experience, poses no problem.
Where do we start with no data team?
Start by checking the reliability of your measurement, before buying any tool. Compare a full month of conversions reported by your ad platforms against real back office orders. Until that gap is explained, any automation will amplify the error instead of correcting it.
The next question to ask yourself
The gap is not widening between companies that use AI and the rest, since access to these tools now costs almost nothing. It is widening between those feeding the tools reliable data and material of their own, and those producing, faster, content nobody can tell apart.
The two concrete workstreams for the next twelve months are therefore modest in appearance: make what you measure reliable, and make what you know and your competitors do not citable. The rest is tooling, and tooling changes in a day.
A third workstream joined those two on 2 August 2026, and it is not up for discussion: knowing which AI systems already run in your business. Chatbot, recommendations, generated product pages, fraud screening: the AI Act applies to your shop, and it all starts with an inventory nobody keeps.
Unsure what AI can genuinely bring to your marketing? On a first engagement we look at three things: the gap between your declared conversions and your real sales, the process costing you the most time, and what your pages say about you when an AI assistant summarises them. Let’s talk.