10 min read
Ad systems on the open web meet most visitors as strangers. Teads gave its click model a short memory of recent browsing, let an LLM suggest training changes from a reviewed list, and reports more clicks at lower cost in a seven-day live test.
TL;DR #
- Ads on the open web often see only a few loose events per person, and nothing at all when someone opts out of tracking.
- Teads, the advertising technology company, turned up to 16 recent events into a compact summary of each visitor and fed it to its click ranker. An LLM, the kind of AI behind ChatGPT, then proposed training changes from about 150 reviewed options.
- In a seven-day test with traffic split 50/50, click-through rate rose 2.13% and visit rate rose 2.37%, while effective cost per click fell 1.13% and cost per visit fell 1.47%.
- Teads ran and published the work, reports 80% confidence intervals and keeps traffic volumes private. The copyable lever is a reviewed menu of allowed changes, with a live control before anyone claims growth.
Picture an ad slot on a news site. A stranger loads the page, and within a few milliseconds the ad system has to decide whether showing that person an ad is worth paying for. Ivan Can Arisoy, a researcher at advertising technology company Teads in Montpellier, works on exactly that moment. A shopping app can remember years of clicks. An ad on the open web may see a handful of disconnected events, or nothing at all.
The usual fallback is a set of counters and buckets such as device, publisher, recency and frequency. They record what happened and lose the order it happened in. Teads trained a compact model on short browsing sequences instead. It reused the result inside its production click ranker, and it kept a stable route for visitors with no history.
Teads researchers posted the paper to arXiv, the open repository for research papers, on 30 July 2026. Click-through rate, the share of ads that get clicked, rose 2.13% in a live test, and effective cost per click fell 1.13%. A second objective built to drive visits after the click produced 2.37% more visits and a 1.47% lower cost per visit. Acquisition teams watch exactly those numbers. The paper reports no advertiser sales or Teads revenue, so the effect on the business stays open.
A short history had to do more work #
The Teads user foundation model starts from a thin record. A foundation model is one trained broadly first and then reused for several jobs. Each event holds a publisher, an advertiser, an interaction type and a timestamp. The production sequence holds no more than 16 events and expires after seven days. That limit was deliberate, because open web advertising AI cannot count on a customer account or on stitching one person together across every browser and device. History exists only for users who opted in. For an opt-out request, the model sees just the current bid-request event.
Teads taught the model to read a browsing sequence #
The encoder, the part of the model that turns a sequence into numbers, learned in two ways. In the first, Teads hid parts of an event sequence and asked the model to rebuild them. In the second, it split one history into an earlier and a later half and taught the model that both halves came from the same anonymous visitor. The summary this produces is called an embedding, a short list of numbers that stands for a whole history. Teads then fine-tuned it on click prediction and attached it to the existing production ranker through a small adapter.
The new model sits beside the bidding decision. A dedicated data service keeps the short history and builds the embedding, and the demand-side platform, the software that places bids for advertisers, uses it next to its usual request features. Teads reports that micro-batching and a one-millisecond maximum wait kept the exchange inside its time budget for each bid.
The embedding carried over to other jobs. In offline tests it improved two different click rankers and a separate model that predicts whether a bid will win. That win-rate model gained 1.197% in relative information gain, a standard score for how much sharper a model’s predictions become, even though the embedding was fine-tuned only on clicks.
LLM model optimization inside a fence #
The idea that travels furthest sits in the tuning loop. An LLM, short for large language model, helped improve the training setup, and Teads did not let it rewrite the model freely. Researchers prepared a catalogue of roughly 150 possible changes, called lifters, each with literature references and a code-level definition. The catalogue covered ten areas, including embeddings, temporal features, objectives, optimisation, regularisation and data augmentation. In every round the LLM saw the approved catalogue, earlier trial results, the current setup and the training time, and it could propose one lifter or a small bundle.
A harness checked each proposal before any computing was spent, so candidates that went over the parameter or training budget were rejected at once. Survivors were trained, tested on data from after the training period, and compared with the running baseline. Accepted changes stayed inspectable and could be replayed in order. Growth Signal covered a looser approach at Kuaishou, where an A/B Agent ran the tests and its first two ideas hurt the platform.
A growth engineering team can borrow this without handing the product to a research agent. Define the allowed changes, attach a cost ceiling and a success metric, keep every trial, and require a live experiment before release. The AI proposes inside the boundary, and the evaluation system decides whether a proposal earns another step.
Across 100 search rounds, 31 candidates joined the best configuration. A full retrain with the chosen setup cut total loss by 25.2% for 2.6% more training time, and the offline click score sharpened by 7.3% over the encoder from before the search. A version without the masked and contrastive stage scored below both, so that stage seems to have added useful signal.
Seven days, half the traffic each #
Offline gains can vanish once a system meets live traffic, so Teads moved the chosen model into a seven-day production A/B test with an even split. This Teads A/B test produced the headline numbers. The control used the existing click ranker, and the treatment used the same ranker plus the new embedding and adapter. Campaign budgets were divided equally between the groups, which matters because Zalando’s ad tests shared one budget and that skewed the results. Visitors with a persistent identifier were assigned by a hash and stayed in one group across requests. Traffic with no identity across requests was randomised bid request by bid request.
On click-optimised traffic, click-through rate rose 2.13%, with an 80% confidence interval from 1.55% to 3.17%, and effective cost per click fell 1.13%, with an interval from 0.64% to 2.07% lower. On visit-optimised traffic, the post-click visit rate rose 2.37%, with an interval from 1.49% to 3.31%, and effective cost per visit fell 1.47%, with an interval from 0.48% to 2.59% lower. An 80% interval is a range the true effect should land in four times out of five. All four company-reported intervals excluded zero.
Where the evidence stops #
The comparison supports a claim about the full deployed treatment. That treatment combined sequence pre-training, architecture choices, the selected optimisation path, serving infrastructure and the adapter, so the LLM’s own contribution cannot be separated out. The paper also compares the LLM proposer with no other search method, such as random search, hill climbing or a person working from the same catalogue and compute budget. The authors present the optimiser as evidence that the workflow can work, and they make no claim that an LLM is the best way to search.
They describe other limits openly. Offline results come from one training seed, so they show no run-to-run variation, and the search picked the best of 100 candidates, which can flatter the winner even though later evaluation used separate data. Teads uses 80% intervals as its internal standard, looser than the 95% common in public research. The paper gives the traffic split and the duration and keeps impressions, advertisers, campaigns, geographies and absolute metric levels private.
How a smaller team can borrow the pattern #
The foundation model needs specialist infrastructure. The operating sequence around it is simpler.
- Represent the smallest useful sequence instead of waiting for a perfect customer history.
- Keep the new representation modular so several downstream decisions can reuse it.
- Give the AI optimiser a reviewed catalogue of permitted changes rather than an open repository.
- Reject candidates automatically when they breach cost, latency or size limits.
- Use time-separated offline data for screening, then require a live control before claiming growth.
- Report the online metric and its uncertainty separately from the optimiser’s offline score.
The European angle is concrete. Teads was founded in 2011 and keeps a major technology centre in Montpellier, where Arisoy works. The experiment takes on privacy-limited histories, a familiar constraint across Europe’s fragmented consent and browser environment. For teams there, privacy-safe ad targeting has to work with incomplete histories, because identity is not always available.
The next test #
Teads has shown that a bounded LLM search can sit inside a production ad model. The next test is whether that search beats a simpler one on the same catalogue and budget, and whether the click gain holds under the stricter 95% interval most public research uses.
Sources and disclosure #
- Building a User Foundation Model for the Open Web , Teads paper posted to arXiv on 30 July 2026
- [Ivan Can Arisoy on LinkedIn](https://fr.linkedin.com/in/ivan-can-arisoy) , Teads researcher in Montpellier
- [Teads office directory](https://www.teads.com/teads-offices/) for the Montpellier technology location
- [Teads investor FAQ](https://investors.teads.com/ir-resources/faqs) for the 2011 founding and 2025 combination with Outbrain
Teads researchers ran and published the experiment, so the figures come from the company. The paper sets out the comparison design, confidence intervals, ablations and limitations. Independent replication has not appeared in the sources reviewed here.
Related reading #
- [How Fyxer used AI to run 360 growth tests with just four engineers](https://industrycontents.com/ai-experimentation-workflow-fyxer/)
- [LinkedIn asked who its ads could actually persuade](https://industrycontents.com/linkedin-causal-incremental-marketing-targeting-experiment/)
- [Some buyers take days to convert. Kwai stopped punishing ads for it](https://industrycontents.com/kwai-ad-system-late-conversions/)
FAQ #
Did the LLM cause the production lift? #
The test measures the effect of the full treatment. Teads did not compare the LLM search with another proposer on the same catalogue and budget, so the LLM’s own contribution is unknown.
How did the system treat opt-out traffic? #
It built a one-event sequence from the current bid request and randomised those requests individually. Offline results stayed positive in the bucket with no history.
Can a smaller team reproduce this? #
A smaller team can reproduce the bounded change catalogue, the automated budget checks, a trial ledger, a time-based evaluation split and a live control. Training and serving the foundation model still need specialist infrastructure.
Compare more AI experiments and their evidence in the Growth Signal Index, browse the Growth Signal archive, or use the benchmark tools in the Industry Contents Lab.