Comment les outils d'espionnage publicitaire capturent les publicités natives (Chaîne d'approvisionnement expliquée)

Imaginez un outil d'espionnage publicitaire et la plupart des gens imaginent un robot qui charge des pages web et capture les publicités qu'il trouve. Cette image est fausse, et l'écart constitue toute l'histoire. Une capture d'écran ne saisit que ce qu'une session de navigateur a reçu à un instant donné, dans un pays. C'est une tranche minuscule et biaisée d'un flux qui tourne des milliers de fois par jour et se personnalise pour chaque visiteur.

Les outils qui construisent des ensembles de données fiables de publicités natives travaillent un niveau en dessous des pixels. Ils parlent le même langage machine‑à‑machine qu'utilise le widget d'un éditeur pour récupérer ses publicités, et ils reconstruisent l'intégralité du chemin depuis l'emplacement sur la page jusqu'à la page de destination de l'annonceur. Cet article parcourt ce chemin de bout en bout, et s'appuie sur l'architecture réelle qu'OpenAdLibrary utilise pour le faire à grande échelle : 589 036 créations, 25 933 annonceurs et 5,4 millions d'observations publicitaires à travers 42 réseaux (index OpenAdLibrary, juin 2026).

Si vous voulez savoir ce qui se cache sous le capot d'un outil d'espionnage publicitaire natif avant de faire confiance à ses données, c'est pour vous.

How do ad spy tools capture native ads?#

Ad spy tools capture native ads by calling the same recommendation API that fills a publisher's widget, then reading the advertiser, headline, image, and click URL straight from the JSON response. They store the creative at full quality, resolve the click URL out-of-band to reach the landing page, and parse the redirect chain to label every intermediary in the supply chain. No live ad ever gets clicked.

That is the whole game in one paragraph. The rest unpacks each step, why the obvious approaches fail, and what separates a thin dataset from a deep one. For the conceptual primer first, start with what a native ad spy tool actually is and how it differs from social-ad libraries.

The native ad supply chain, hop by hop#

You can't capture something you don't understand. A native ad is not a static image dropped into a page. It is the visible tip of a supply chain with several moving parts. Here is the path a single Taboola or Outbrain unit travels:

1. The publisher page loads a native ad widget. That is the "Around the Web" or "You May Like" box. At first paint it is mostly an empty container.
2. The widget script calls a recommendation endpoint. This is a JSON API on the demand platform's domain. It passes the publisher ID, the slot, the user's approximate geo, and device signals.
3. The demand platform runs an auction. In programmatic native, advertisers bid for that impression in milliseconds, and the winners come back in the response.
4. The response returns the ads as structured data. Each item carries a title, a thumbnail image URL, a branding or advertiser name, and a click‑tracking URL.
5. The click URL is a redirect, not a destination. Click it and you bounce through one or more tracker and exchange domains before you reach the advertiser's real page, which is frequently a pre‑lander or advertorial rather than a product page.

The single most useful realization in native ad intelligence: the ad data you want already arrives as clean JSON before a single pixel is painted. Screenshotting the rendered widget is reverse‑engineering something you could have just read.

Every step is a capture opportunity and a classification problem at the same time. Steps 2 and 4 hand you the creative and the advertiser. Step 5 hands you the supply chain and the landing page. Tools differ wildly in how many of these they actually use.

API-only harvesting vs browser screenshotting#

There are two fundamentally different ways to harvest native ads, and the choice cascades into cost, coverage, and data quality.

The browser route is the obvious one, and plenty of legacy tools started there. It is also expensive and shallow. Spinning up real Chromium for every page is heavy, slow, and easy to fingerprint and block. Worse, a render only ever shows you the few ads served to that one session. You are sampling a slot that personalizes and rotates thousands of times a day and calling it a dataset.

The API-only route is what OpenAdLibrary's native harvester runs on. Instead of rendering pages, it requests the same recommendation feeds the widgets call, across many geographies and rotating device and identity personas, and reads the ads directly. In practice that is roughly an order of magnitude cheaper per ad than driving a browser, which is exactly why it can run continuously and catch far more of the rotation. No browser also means no clicking live ads. The data comes off the feed, not off a rendered impression.

That difference shows up in the numbers. On Taboola alone the index holds 157,727 creatives, and on Outbrain 84,252 (OpenAdLibrary, June 2026). You do not get to six figures per network by screenshotting pages one at a time. Here is a live Taboola finance ad pulled straight from the feed, not a render:

This is the difference between a tool that shows you a handful of a competitor's ads and one that shows you the full spread. For affiliates and media buyers, the spread is the whole point. See how affiliate marketers use native spy tools to find rotations worth modelling.

Capturing the creative at full quality#

Once the feed is parsed, the creative image URL points to the advertiser's original asset on a CDN. A good harvester fetches that asset directly and stores it at native resolution instead of keeping a downscaled screenshot crop.

Why it matters in practice:

• Reverse image search and dedup only work on the original asset. Cropped screenshots break perceptual hashing and bloat your dataset with near‑duplicates.
• Creative analysis needs the real pixels. The hook, the face, the text overlay, the color treatment. You cannot study what made a winner win from a thumbnail.
• Cross‑network asset reuse detection depends on the source file. Matching the same image running on Taboola, MGID, and Revcontent at once is only reliable when you hold the original.

Storing the original also means the creative survives after the campaign ends and the CDN URL 404s, which is most of why an ad library has value at all. Across the index, OpenAdLibrary has held onto 926,259 landing‑page captures tied to those creatives. The campaign dies; the record does not.

The health vertical is where this pays off hardest, because the creatives are aggressive and they recycle constantly. Health is the third‑largest vertical in the index at 14,895 creatives, behind finance (17,232) and insurance (15,629). Here is one running on Taboola for 26 days straight:

Following the click to the landing page (without clicking)#

This is the step most tools skip, and it is where the real intelligence lives. The click URL in the feed is a tracking redirect. The destination behind it, the landing page or pre‑lander and the advertiser whose domain it sits on, answers the only question that matters: who is actually running this, and where are they sending traffic?

The naive approach is to fire the click in a browser. That can register as a billable click on a live impression and burn budget you do not own. Not acceptable. The correct approach is to resolve the redirect chain out‑of‑band: replay the hops server‑side, usually from a clean residential exit in the relevant geo, to retrieve the final URL and capture the landing page itself, without ever triggering a live billable click.

Done well, this surfaces three things at once:

• The real advertiser behind a generic‑looking branding name in the feed.
• The pre‑lander or advertorial, the bridge page native buyers rely on, which never appears in the widget.
• The geo‑gated destination. The same ad often routes to different landers by country, and only a multi‑geo resolver catches that. The example below is geo‑targeted at Australia, which you would miss entirely capturing from a US exit:

OpenAdLibrary traces each click to its landing page this way and stores the destination next to the creative, so a creative is never an orphaned image. It is tied to the advertiser and the funnel it feeds.

Classifying the ad‑tech supply chain#

Capturing the hops is one thing. Making them legible is another. Between the widget and the lander, a single ad can pass through a demand platform, an exchange, several tracker domains, and a redirect service. Classification is the work of parsing that chain and labelling each node.

A capable system keeps a dynamic registry of known networks, trackers, and redirect domains, then matches each hop against it to answer:

• Which demand platform served the ad (Taboola, Outbrain, MGID, Revcontent, MediaGo, Yahoo, Microsoft Audience Network)?
• Which trackers and exchanges sit in the redirect chain?
• Who is the end advertiser, normalized across the many aliases de marque que le même acheteur cache derrière?

Hard‑coding tracker lists is a losing game. The domains rotate constantly, so it turns into whack‑a‑mole. A registry that updates as new patterns appear is the only approach that holds up. Get the supply chain right and you can answer questions a flat ad list never can: which advertisers concentrate on which networks, which trackers signal a particular affiliate stack, which redirect services correlate with the most aggressive offers.

The pattern is visible in the data. Both Taboola and Outbrain are dominated by the same three verticals, but the order flips. On Taboola the lead is health (6,048 creatives), then finance (5,558) and insurance (4,303). On Outbrain finance leads (2,640), then insurance (2,615), then health (2,016) (OpenAdLibrary, June 2026). That kind of per‑network advertiser concentration is exactly what supply‑chain classification exists to expose.

If the terminology is new, the glossary entry on the ad supply chain lays out the roles, and programmatic native advertising explains the auction that decides which ads you capture in the first place.

Coverage, geo, and personas: why one capture isn't enough#

A native slot is personalized. The ads it serves depend on geo, device, time of day, and interest inferred. So a single capture, from one IP, un profil d'appareil, un instant, est un échantillon biaisé d'une rotation bien plus large.

Serious harvesting treats this as a sampling problem:

• Geo rotation. The same widget on the same publisher serves different advertisers in the US, UK, AU, and DE. One geo gives you one tranche.
• Device and identity rotation. Desktop, Android, and iOS sessions, plus rotated personas, surface different demand and different creative formats.
• Cadence. Rotations turn over throughout the day, so capture must be continue, not a crawl ponctuel.

This is also why la longévité et la diffusion sont les signaux de gagnant les plus honnêtes disponibles sans données internes. Vous ne pouvez pas voir les dépenses d'un concurrent, mais vous pouvez voir combien de temps une création a fonctionné et à quel point elle s'est répandue à travers les éditeurs, les zones géographiques et les réseaux. Actuellement les créations observées en continu les plus longues de l'index tiennent 28 jours d'observation ininterrompue. SmartAsset a diffusé "Ask a Pro: How Can I Avoid Paying Taxes on IRA Withdrawals?" sur Outbrain pendant les 28 jours complets :

Worth being precise here: 28 days is the span of our observation window, not a claim that the creative has run for exactly 28 days and no more. The industry lore about 90‑day winners is a separate thing, useful as a rule of thumb but not something our index measures. What we can stand behind is the observed run: an ad that has been live across dozens of placements for weeks is telling you something a spend number cannot.

What this means for the data you actually get#

Pull the threads together and the harvesting method directly decides what you can do with the output.

A tool that only screenshots gives you the first row. A tool that harvests feeds, traces clicks, and classifies the supply chain gives you all of them. That is the difference between a curiosity and a research instrument. On top of clean capture, OpenAdLibrary layers tooling that uses it: Creative Studio to remix what works, Optimize to act on it, Copy DNA to break down the angle, plus a full API and MCP so you can pull the data into your own stack.

The transparency tailwind#

For anyone weighing the legitimacy de tout cela, la direction réglementaire tend vers plus de données publicitaires publiques, pas moins. Le Digital Services Act de l'UE exige désormais que les très grandes plateformes (Meta, TikTok, Google, et similaires) maintiennent des dépôts publics et interrogeables listant l'annonceur, le payeur et les dates de diffusion pour chaque annonce. Les widgets de découverte native ne sont pas couverts par ce mandat, mais le principe est établi : les annonces servies au public sont de plus en plus traitées comme information publique. Les outils qui ne capturent que les publicités publiques, et ne cliquent jamais sur des unités en direct de façon facturable, s'inscrivent confortablement dans cette tendance. (Contexte général, pas un avis juridique. Vérifiez les conditions de chaque plateforme pour votre usage spécifique.)

The bottom line#

Stripped to essentials, here is how the good ad spy tools work. They read the recommendation feed instead of screenshotting the page. They store the original creative instead of a crop. They resolve the click out‑of‑band to reach the landing page instead of firing a billable click. And they classify every hop in the supply chain instead of leaving you a flat list. Coverage comes from rotating geos, devices, and personas continuously, and the most trustworthy winner signal is longevity and spread.

When you compare vendors, those are the questions to ask. Our tested, ranked breakdown of the best native ad spy tools scores each one on exactly these axes. If budget is the constraint, there is also a guide to running native ad research for free.

OpenAdLibrary built its native harvester on every principle above: API‑only capture, full‑quality creatives, click‑traced landing pages, and a live supply‑chain registry. It is open and affordable rather than $80 to $400 a month. Start free and browse 200 ads, no credit card required.