Two businesses in the same trade, the same monthly budget, the same area. One brings in enquiries, the other burns its budget. The difference almost never comes from the campaign settings — it comes from what the system understood, or failed to understand, about what each one sells. And over the past two years, the way it understands has changed profoundly.
Good news: it isn't a total black box. Meta publishes how its advertising models work on its engineering blog. It's technical, it's in English, and it isn't written for advertisers — but it's the most reliable source there is, far more than advice passed around by word of mouth. This article translates the essentials into plain language.
What happens between your ad and someone's screen
In July 2026, Meta published the description of a component called Hierarchical Interest Representation — literally "hierarchical representation of interests". It's the layer that manufactures the understanding of people and products, ahead of everything else. It works in five floors.
What it looks at to begin with
Users · Advertisers · Products · World KnowledgeFour raw materials: people, advertisers, products — and a fourth Meta calls "world knowledge". Concretely, your copy, your images and your videos run through a language model to understand what the advertised thing is.
In plain terms: the machine no longer just counts clicks. It looks at your ad and tries to understand what you sell.The map of connections
Enriched Engagement GraphA vast network of relationships between billions of people, advertisers and products. The links are typed, weighted, and lose value over time.
In plain terms: a map of "who is interested in what", where yesterday counts for more than last year.Understanding by levels
Hierarchical EncoderThe heart of the system. It doesn't use off-the-shelf labels like "interested in gardening": it builds its own notions of interest, at several levels of detail.
In plain terms: Meta no longer has the same definition of an "interest" as the dropdown menu you see in the ads manager.The result: a fingerprint and meaning-words
Universal Embeddings · Bag-of-Meaning TokensEvery person and every ad comes out described two ways: a numerical fingerprint, and a small bag of "units of meaning" the system can look up in an index — and that humans can read back.
In plain terms: your business ends up summarised in a handful of notions. Those are what decide who you're shown to.What uses it next
Personalization · Retrieval · Supervision · Model ArchitectureThis material feeds the next floors: the selection of which ads enter the running, the central model that predicts, and the final ranking that decides which ad wins the impression.
In plain terms: the whole rest of the chain works from that understanding. If it's wrong, no campaign setting makes up for it.A word on the stated goal, because it explains everything else. Meta writes that feedback "at the bottom of the funnel" is sparse — in other words: clicks are plentiful, purchases are rare. All this machinery exists to find people with a real, lasting interest, not people who press a button. That's good news for an SME selling kitchens, and less good news for anyone who was buying cheap clicks.
Meta no longer uses your "interests" boxes the way you think
This is the point that surprises advertisers most. The system builds its own notions of interest, at several levels of precision, and the rule it applies fits into one sentence from the publication: dense, stable links correspond to a broad level, rare and specific links to a fine level.
Direct consequence: carefully slicing audiences by interest has less and less effect, because the system's internal map is richer than the list it offers you. That doesn't mean you should specify nothing — age, language and geographic area remain real and useful constraints, especially in French-speaking Switzerland. It means spending an hour stacking ten interests is no longer worth that hour.
Your creative has become your targeting
This is the most important consequence, and the one that demands the biggest change of habit.
Since the system reads your image, your video and your copy to understand what you're offering, what you show decides who you see. A blurry photo taken in a hurry with the caption "get in touch" says nothing: the machine doesn't know whether you lay tiles, sell tiles, or renovate whole flats. So it can't place you anywhere with confidence.
And there's a second, less obvious consequence. If each ad is understood for what it says, then every genuinely different angle is a different door to a different group of people. Twenty variants of the same visual with a different-coloured button say the same thing twenty times: that's a single door.
- The same visual in 20 versions
- Changing the button, the colour, the crop
- Stacking 10 interests in the targeting
- Cutting and relaunching every two days
- Targeting done by hand, the creative is decorative
- 4 or 5 angles saying distinct things
- One "urgency" angle, one "price", one "before/after", one "who we are"
- Broad targeting, bounded by language and area
- Letting it run long enough to learn
- The creative carries the targeting; the settings only frame it
For a Swiss SME, "four or five angles" is hardly a production undertaking. For a landscaper: a before/after of a terrace, a close-up of work done well, a client speaking for twenty seconds, a seasonal visual. For a beauty salon: the result, the premises, the team, a dated offer. The cost isn't in the camera, it's in deciding four different things to say.
What the machine expects from you in return
This point is the least spectacular and the most decisive. All this machinery is looking for people who buy, not people who click. So it needs to be told what happened after the click.
If your site reports nothing back — no tracking of enquiries sent, no distinction between a form filled in by a genuine prospect and a visitor who bounces — the system optimises towards the only thing it can see: the click. You then get exactly what you asked for, namely clicks, and conclude that "advertising doesn't work". It worked very well: it chased the wrong goal, because it was the only one available.
This is plumbing, not creativity: wiring up conversion tracking properly, distinguishing a serious enquiry from a simple contact, and where possible feeding back what becomes of the enquiry once it's in your CRM. Without that, you're driving a powerful engine with the gauge disconnected.
The most useful question to ask before increasing a budget: "do I know, of the last 30 enquiries received, which one became a client?" If the answer is no, extra money won't improve anything.
Get my free audit →The figures Meta publishes — and the ones circulating wrongly
Since money is involved, better to cite what's verifiable and set the rest aside.
What Meta reports. Its central model, called GEM, reportedly brought "a 5% increase in ad conversions on Instagram and 3% on the Facebook feed" in the second quarter. Its retrieval engine, Andromeda, reduces "tens of millions" of candidate ads to a few thousand before the final ranking, with "+6% recall" and "+8% ad quality on certain segments".
What needs qualifying. These are Meta's figures on its own products, measured its own way. They indicate a direction, not a promise of results for your campaign. A gain of a few percent at Meta's scale doesn't translate mechanically to a tradesperson in Prilly.
Two widespread claims we don't use. The first: "GEM is four times more effective". That's a distortion — the factor of four published by Meta concerns training compute multiplied by four over twelve months, not the performance of your ads. The second: "broad targeting delivers 49% more return on investment". We found no primary study behind that figure, only articles citing each other. A figure you can't trace back to its source has no place in a budget decision.
When Meta Ads isn't your channel
It has to be said, because it's true for a good share of the businesses we meet.
Advertising on Facebook and Instagram addresses people who weren't looking for anything. It creates desire or reminds people you exist. If your clients arrive typing "emergency heating repair Lausanne" at 6am because they're cold, that isn't your ground: those are won on the map and in search results, and we explain it in detail in the article on the three paths your clients take.
Meta Ads becomes relevant when the decision is made with the eyes or can be prompted: a restaurant launching a seasonal menu, a salon opening slots, a landscaper showing a before/after, a shop with a dated offer. And it becomes frankly counterproductive when nothing is ready behind it: sending paid traffic to a site where you can neither book nor easily write is paying for people to visit a closed door.
Our real work: refusing to launch too early
At Renova Softwork, we regularly refuse to start a campaign on the day we're asked to. Not out of excessive caution, but because a campaign launched on a shaky base reveals nothing: it spends a budget and leaves no learning behind.
The order we apply is simple. First check that the click has somewhere to land and something to do there. Then wire up the measurement, so you know what becomes of an enquiry. Only then produce four or five creatives that say different things, and let it run long enough for the system to learn something. That last point is the one advertisers bear least well: cutting a campaign after three days means erasing the learning and starting from zero every time.
And we'd rather say "your advertising budget is premature" than take it. It's the same principle as with chatbots : a powerful tool placed on a fragile base amplifies the problem instead of solving it. That's the point of our offers for digital marketing for SMEs — start with what holds, and only open the tap afterwards.
Frequently asked questions
Should you still pick interests in the targeting?
Less and less, and that follows from the workings Meta describes: the system builds its own notions of interest, at several levels of detail, and its internal map is richer than the list offered in the interface. The constraints that remain genuinely useful are language, age and geographic area — particularly in French-speaking Switzerland, where language is a decisive filter. Stacking ten interests is no longer worth the time spent on it.
How many different ads should you produce?
Few, but genuinely different. Since the system reads the content of each ad to understand what it says, twenty variants of the same visual say the same thing twenty times. Four or five distinct angles — the result, the price, before/after, the team, a dated offer — open four or five different doors. The effort goes into choosing the messages, not into the number of files.
Why does my campaign bring clicks but no clients?
That's the classic symptom of incomplete measurement. The system optimises towards what you show it: if your site doesn't report back what happened after the click — a serious enquiry, a booked appointment, a sale — the only thing it sees is the click, and it will keep bringing you more. Wiring up conversion tracking properly, and distinguishing a real enquiry from a passing visit, changes results more than any targeting setting.
Should you switch off an ad that isn't performing after two days?
That's one of the most expensive reflexes. These systems rely on patterns built over time, and cutting or changing a campaign every two days erases the learning in progress and sends it back to zero. Better to launch fewer campaigns, with genuinely distinct creatives, and give them time to produce usable data.
Are Meta's automatic tools worth it for a small business?
They go in the system's direction and deserve testing, provided you give them material: several genuinely different visuals and texts, not one image in variations. Meta for its part reports gains for advertisers using its automated formats, but those are its own figures, measured its own way, and they say nothing about the result in your sector. These tools amplify what you give them: on a weak base, they amplify the weakness.
These algorithms change all the time. Will this article still be valid in a year?
The details will change, the direction much less. For several years now, every Meta publication has pointed the same way: fewer manual settings on the advertiser's side, more automatic understanding of ad content, and optimisation aimed at the purchase rather than the click. The practical consequences — look after your creatives, produce genuinely different ones, measure what happens after the click, allow time to learn — remain valid even if the model names change.