Ad fraud & invalid-traffic (click fraud) detection

Part of every advertising budget is spent on clicks, impressions and installs that no real, interested person ever produced: bots, click farms, hidden ad stacking, spoofed domains, competitor clicking and fake mobile attributions. Detection combines bot and device signals, supply-chain checks and anomaly or supervised models to filter that traffic before or after payment, exclude bad sources and document claims for refunds. Independent vendors measure invalid traffic in the high single to double-digit percent of paid traffic, so even small advertisers can recover real budget.

Problem & value

• Who has it: SMBs buying search and social ads (click fraud on Google/Meta campaigns), e-commerce and lead-gen advertisers, app publishers paying per install, agencies and programmatic buyers, and publishers whose inventory is impersonated or whose traffic is inflated by purchased visits.
• Today: advertisers trust platform-side filtering and manual checks of suspicious placements; mobile teams use attribution tools; programmatic buyers use pre-bid filtering by verification vendors. Gaps remain: smaller advertisers rarely see the raw data, and fraud adapts to the filters.
• AI value: behavioural and device-level classification at click time, graph clustering of sources that share devices, IPs and timing, and anomaly detection on funnel metrics (click-to-install time, CTR vs. conversion, bounce patterns). LLMs help with evidence summaries and dispute letters, not with the detection itself.
• Scale (vendor-reported, treat as indicative): Lunio's 2026 report puts the 2025 average invalid-traffic rate at 8.51 % across channels, implying about USD 63 bn of roughly USD 740 bn global ad spend; the same report lists Western Europe at 6.34 % [1]. Fraudlogix reports 20.64 % across 105.7 bn impressions (programmatic impressions, a different base) [2]. Other vendors quote far higher totals; the numbers differ because vendors use different definitions and data, and all of them sell detection products. No neutral census exists.

Variants

• Search/social click fraud (SMB): repeated clicks by bots or competitors on paid search; remedy is IP/audience exclusion and claims to the platform.
• Programmatic / display and CTV: invalid impressions via domain spoofing, ad stacking, data-centre traffic; handled by pre-bid verification, allow-lists and supply-path optimisation using ads.txt / sellers.json [3].
• Mobile app install and attribution fraud: click injection, click flooding, SDK spoofing, device farms, fake installs; attribution tools flag them by timing, signature and device-reputation checks.
• Lead-gen and affiliate fraud: fake form fills, cookie stuffing; combine with lead scoring and ANOM-004.
• Publisher side: detect purchased or bot traffic mixed into own inventory to protect CPMs and avoid being delisted.
• Industry vocabulary: the Media Rating Council distinguishes general invalid traffic (known bots, data-centre traffic; filtered by list) from sophisticated invalid traffic (hijacked devices, human-mimicking bots; needs analytics). Use these terms in contracts.
• Private use: not applicable.

How it works

1. event capture (ad-platform logs, tag / server-side events, SDK / attribution postbacks, bid requests)
2. enrichment (IP / ASN / data-centre / VPN reputation, device & browser fingerprint, geo & language consistency, supply-chain lookup)
3. feature building (session behaviour, click-to-conversion time, per-source rates, device/IP graph)
4. scoring: rules + gradient-boosted models (known patterns) + anomaly / graph models (new campaigns)
5. decision (allow / flag / block / exclude placement or IP / mark conversion invalid)
6. feedback (clean conversions to ad platform, exclusion lists via API, evidence pack for refund claim)
7. analyst review and model retraining on confirmed outcomes

• Key components: tabular ML for the score; unsupervised anomaly detection and time-series baselines for source-level spikes; graph clustering to expose farms sharing infrastructure; a small text classifier or LLM for referrer, search-term and form-spam classification.
• Human in the loop: analysts approve new blocking rules, review borderline sources, and decide on refund claims. Hard blocking of real customers is a business risk, so start with flagging.

Data preparation

The relevant data sits in several places: ad-platform logs (Google Ads, Meta), web analytics and server logs with IP, user agent and device data, mobile attribution postbacks, and CRM or shop systems that show which leads actually became customers or were refunded. In an SMB these are usually separate tools without a common key, and exports are often manual CSV files.

Typical gaps: no link between click ID and later sale; lead outcomes not recorded in the CRM; short retention of server logs; consent banners that suppress tracking; and no list of known good traffic sources.

Preparation steps:

1. Inventory the sources and the questions you want answered (PREP-001); get data out of ad platforms and CRM via exports or connectors (PREP-005).
2. Capture click IDs (gclid, fbclid) in forms and store them with the lead; record lead quality and refunds as labels (PREP-020).
3. Tidy spreadsheets into one machine-readable table with consistent campaign naming (PREP-011).
4. Check legal basis for log and IP processing, and shorten or pseudonymise IPs where possible (PREP-022, PREP-025).

Tools: server-side tagging, a spreadsheet or small database, n8n for exports. Rough share of effort: about 25 percent on the buy path, 40 percent when you build your own models.

Implementation paths

Which path for whom

• smb: A. Mature click-fraud and traffic-quality tools exist at SMB prices, many with EU hosting, and the problem is standard for any advertiser.
• enterprise: B. Large advertisers combine their own conversion and CRM data with cloud APIs, as volume makes per-click vendor pricing high and own signals are decisive.
• developer: C. Open-source ML and streaming stacks on own hardware allow transparent detection logic and keep traffic data in-house.

Path A — Buy (recommended for most SMBs)

• Click-fraud protection for search/social SMBs: ClickCease, Lunio, PPC Protect, TrafficGuard, Fraud Blocker; they exclude IPs and audiences automatically via ad-platform APIs.
• Verification for enterprise/programmatic: DoubleVerify, Integral Ad Science (IAS), HUMAN Security, Pixalate, Anura, Fraudlogix.
• Mobile attribution suites: Adjust, AppsFlyer, Singular, Branch, Kochava all offer fraud filters in their attribution products.
• Platforms themselves run invalid-traffic filtering and can credit advertisers; always review the platform's invalid-click reporting first.
• EU/German options: the market is dominated by US/UK vendors; Adjust is Berlin-based. Check each vendor's EU data residency and data-processing agreement; other vendor locations should be verified in procurement.
• Choose A when: spend is moderate, you lack raw data access, and a tag or API integration suffices.

Path B — Build on cloud APIs

• Log pipeline on EU-region cloud: server-side tagging or log export (e.g. Google Ads / GA4 BigQuery export), streaming into BigQuery, Azure or AWS EU regions, plus managed ML (Vertex AI, Azure ML, SageMaker) for gradient-boosted models and anomaly detection.
• Enrichment via IP-reputation and ASN data sources; bot signals via edge providers (Cloudflare Bot Management, Akamai, DataDome (French), HUMAN).
• An LLM API (e.g. Claude, Mistral) can summarise flagged clusters and draft credit claims, with PII minimised.
• Choose B when: you have in-house data engineering, need custom attribution logic or want to combine fraud with ANOM-004 signals.

Path C — Self-host (open source)

• Stack: Kafka or Redpanda, ClickHouse or PostgreSQL, Python with scikit-learn, XGBoost / LightGBM, PyOD or Isolation Forest for anomalies, NetworkX / Neo4j for device-IP graphs, Prophet or statsmodels for baselines; IP data from public ASN and data-centre lists; n8n for alert workflows.
• Hardware: HW1; CPU is sufficient. A GPU is only needed for optional LLM summaries.
• Limit: open source supplies the models and plumbing, not the threat intelligence (device reputation, known botnets) that vendors maintain; expect weaker detection of sophisticated invalid traffic.

Path D — Custom training

Worthwhile for large app publishers, ad networks and exchanges with their own click and install logs. Train supervised models on labelled outcomes (refunded, charged back, zero retention, confirmed bot) and sequence models on behavioural telemetry. Labels are the bottleneck: fraud is only partially observed, so use semi-supervised methods and validate on holdout periods.

State of the art (2026)

• Arms race: vendors report that AI-assisted bots mimic human behaviour better and that sophisticated invalid traffic is rising [2]; one 2026 vendor brief cited by eMarketer claims invalid-traffic rates up to 40 % by mid-2026, a claim from a single vendor that I could not independently verify [4].
• Techniques: ensemble of cheap rules, supervised gradient boosting and behavioural/device telemetry remains the production core; graph methods expose coordinated farms; time-series baselines catch sudden source-level spikes. LLM agents are used in analyst tooling and investigation, not as the primary detector.
• Supply-chain transparency: ads.txt, app-ads.txt and sellers.json let buyers check whether a seller is authorised and cut counterfeit-inventory fraud [3]; they do not detect bots.
• Privacy shift: less third-party-cookie and device-ID data pushes detection to server-side and first-party signals.
• Limits: no ground truth, so vendor "fraud rates" are not comparable; false positives block real customers; sophisticated fraud passes pre-bid filters and is only found post-bid; platform and vendor models are black boxes and credits are often capped by policy.

Shared resources & combinations

• Same building blocks: tabular-ml + anomaly-detection + graph-ml + streaming as in ANOM-001 (payment fraud), ANOM-004 (platform abuse, bot defence), ANOM-033 (returns/POS) and ANOM-036 (game cheating). A shared device/IP reputation store and event pipeline serves all of them on HW1.
• Upstream/downstream: REC-007 (ad optimisation and bidding) should consume cleaned conversion data, otherwise it optimises toward fraud. ANOM-010 (fake reviews) and ANOM-028 (counterfeit listings) use the same web-scraping and graph tooling. SEC-005 (UEBA) and SEC-010 (guardrails against agentic traffic) share bot-signal infrastructure.
• Data pipeline: server-side event collection and CRM outcomes are reusable for churn and lead-scoring models.

Effort, cost & KPIs

• Path A: days to two weeks (tag plus ad-account connection); pricing typically scales with monthly ad spend or click volume.
• Path B: 2-4 months for a pipeline with models and dashboards.
• Path C: 2-4 months; engineering time is the main cost.
• Path D: 6+ months incl. labelling and validation.
• KPIs: invalid-traffic rate by channel and source, budget recovered through credits and exclusions, cost per qualified lead / acquisition before vs. after, false-positive rate (real conversions blocked), share of SIVT detected post-bid, time-to-exclusion of a bad placement.

People & change

Who and when. Management sets the budget goal (recovered ad spend). The marketing lead owns the process and picks the pilot campaigns. IT or the web agency implements tracking and log access, and the DPO is consulted before IP and device data are processed. Suppliers such as ad agencies and platforms must agree to evidence sharing for credit claims.

Works council. No in the typical case: the system evaluates external traffic, not employee behaviour. Check only if the tool also logs how marketing staff work, in which case §87(1) no. 6 BetrVG may apply; §90(1) no. 3 asks for timely information when AI changes work processes.

Data protection. IP addresses, device fingerprints and click IDs are personal data. Conclude an AVV (Art. 28 GDPR) with the vendor, document the legitimate interest in fraud prevention, and run a DPIA only if profiling is extensive (GUIDE-005).

AI literacy and control. Show marketers what a fraud score means and where false positives arise, so real customers are not blocked (GUIDE-006). Start with one or two campaigns in a pilot, review exclusion lists by hand before applying them, and compare results against the baseline (GUIDE-009). A typical concern is "the tool blocks real buyers"; address it with a monthly review of excluded sources.

Risks, regulation & compliance

• EU AI Act: minimal risk. It is a B2B integrity tool without an Annex III use; AI literacy (Art. 4) applies. Not biometric categorisation or profiling of individuals for access to essential services. Caveat: if blocking decisions lock real users out of services (e.g. a credit offer), review it under the relevant sector rules.
• GDPR / ePrivacy: IP addresses, device fingerprints and behavioural telemetry are personal data. Legitimate interest in fraud prevention (Art. 6(1)(f), recital 47) is commonly relied on, but needs a documented balancing test and data minimisation. Storing or reading device information on the user's terminal also falls under Art. 5(3) ePrivacy (in Germany § 25 TDDDG) and may need consent unless strictly necessary. Check that the vendor is a processor with a DPA and EU or adequately protected transfers.
• Contracts: platform credit claims have deadlines and evidence requirements; keep logs. Vendor "guarantees" often exclude SIVT.
• Risks and mitigations: false positives (start with flag-only, review samples), vendor lock-in and conflict of interest (vendors paid on spend), adversarial adaptation (retrain regularly), mismeasurement (compare to CRM outcomes, not just platform metrics).

Getting started

1. Pull click and conversion data for the last 90 days and compute simple indicators: bounce rate and conversion rate by placement, geography and device, plus bursts of clicks from one IP or ASN.
2. Enable the platform's invalid-click reporting and review credits; add obvious exclusions (data-centre ASNs, repeated IPs).
3. Trial one or two Path A tools on a subset of spend for 4 weeks; compare against CRM lead quality, not only against the vendor's dashboard.
4. Feed cleaned conversions back to ad platforms and bidding (REC-007).
5. Document the legal basis, consent handling and DPA before scaling.

Sources

1. Piracy Monitor — Invalid ad traffic for 2025 estimated at $63B globally (reporting Lunio's Global Invalid Traffic Report 2026) — https://piracymonitor.org/invalid-ad-traffic-for-2025-estimated-at-63b-globally-ai-increasingly-a-driver/
2. Fraudlogix — 2026 industry report and ad-fraud statistics — https://www.fraudlogix.com/stats/ad-fraud-statistics-2026
3. IAB Tech Lab — ads.txt / app-ads.txt (Authorized Digital Sellers) — https://iabtechlab.com/ads-txt/
4. eMarketer — AI-powered ad fraud distorts campaign data as invalid traffic hits 40% — https://www.emarketer.com/content/ai-powered-ad-fraud-distorts-campaign-data
5. IAB Tech Lab — OpenRTB specification repository — https://github.com/InteractiveAdvertisingBureau/openrtb
6. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj