Facebook Attribution Settings Optimization: The Algorithm’s Secret Lever
13 min read
A contaminated conversion signal, a bot purchase, a fraudulent lead form fill, entered your attribution window and the algorithm studied it.
Simul Sarker
Founder & Product Designer of DataCops
Last Updated
May 28, 2026
Project Andromeda fully deployed in October 2025. It is the AI running Meta Advantage+. And it changed the cost of bad attribution data without changing anything on your attribution settings page.
The old Meta delivery algorithm was slow to learn. A contaminated conversion signal, a bot purchase, a fraudulent lead form fill, entered your attribution window and the algorithm studied it. Over days, sometimes weeks, it built targeting patterns from that data. The contamination caused drift. It was expensive. It was also gradual enough that most advertisers never diagnosed it cleanly.
Andromeda acts on signals within hours. The same contaminated conversion that used to degrade your targeting slowly now reshapes your audience model by the next morning. The attribution window settings are identical. The mechanism is identical. The speed of the damage is an order of magnitude faster.
Every attribution guide still treats the window as a reporting preference. Pick 7-day click or 1-day view, watch your numbers, adjust. That framing was incomplete before Andromeda. It is actively misleading now.
The attribution window is not a reporting setting. It is the boundary that defines which conversions Meta's algorithm studies to decide who sees your ads next. Change the window and you change the training picture. Feed contaminated events into that window and you train a fast, powerful algorithm on a lie. Andromeda acts on that lie within hours.
This guide covers what the window actually does, what each setting is actually for, and the architecture that determines whether the events inside your window are worth training on.
Quick answers
What is the best Facebook attribution window?
For most ecommerce: 7-day click, 1-day view. It captures realistic consideration without over-crediting weak view-through signals. For lead generation with a multi-day sales cycle: 7-day click, no view-through, because view-through attribution for B2B inflates volume and degrades lead quality signals. The window selection matters. It matters less than whether the conversions inside it are from real humans.
How does the attribution window affect ad delivery?
Directly, and this is the part every guide understates. The window defines which conversions get credited to which ad impressions. That credited set is what Andromeda studies to build audience targeting profiles. A wider window gives the algorithm more examples to learn from. It also gives it more contaminated examples if your conversion events are not filtered before they fire. The dropdown is wired to delivery, not just reporting.
What is the difference between click-through and view-through attribution?
Click-through credits a conversion when the user clicked the ad before converting. View-through credits it when the user only saw the ad and converted later without clicking. View-through is a softer signal and easier to inflate: a bot impression that does not click but appears in the same session window as a conversion event gets credited. For Advantage+ campaigns where Meta controls placement, view-through attribution on Instagram and Audience Network at 38% and 67% IVT respectively is pulling significant bot signals into your training data.
Is 7-day click or 1-day click better?
7-day click gives the algorithm more conversions to learn from, which helps smaller accounts exit the learning phase. 1-day click is cleaner but starves low-volume accounts of signal. The practical choice depends on your conversion volume. The more important choice is whether those clicks represent real human intent.
How does the Conversions API improve attribution accuracy?
CAPI moves conversion events server-side so they survive ad blockers and iOS Safari ITP that would otherwise drop 25-35% of pixel events. More real conversions reach the attribution window. CAPI does not filter whether those events are real. It recovers the missing conversions and delivers the fraudulent ones with equal efficiency. CAPI is a more reliable pipe, not a cleaner one.
Why do my Meta Ads keep targeting the wrong audience even with correct attribution settings?
The settings are not the problem. The events inside the window are. If your attribution window contains bot conversions, Andromeda has profiled the traffic patterns behind those sessions and is now targeting audiences that resemble them. The algorithm is doing exactly what you asked it to do. It was given the wrong success definition.
Does changing the attribution window affect ad delivery?
Yes. A different window credits different conversions to different impressions. A different credited set trains a different audience model. The algorithm shifts who it targets based on what the window tells it is a successful outcome. This is not a delayed effect. With Andromeda, the shift begins within hours of the window change.
The mechanism, unambiguous
Meta's Advantage+ algorithm learns from outcomes. You pick a conversion event: Purchase, Lead, Complete Registration. The attribution window determines which occurrences of that event get tied back to which ad impressions.
That tied-together set of impression plus credited conversion is the training data. Andromeda studies it, builds a statistical model of who converts, and pushes your budget toward audiences who match that profile.
Your attribution window settings determine the size and shape of that training set. Your data quality determines whether the training set is real.
Two failures corrupt the training set simultaneously.
First: 25-35% of real human conversion events never reach the window. Your pixel fires from the browser. uBlock Origin blocks it. Brave Shields blocks it. iOS Safari with ITP downgrades the cookie and strips the fbclid. That real purchase from a customer who runs a privacy browser is invisible to your attribution window. Andromeda never learns from your best customers because it never saw them convert. It learns from whoever was left after the blockers ran.
Second: of the events that did reach the window, a significant share came from non-human sessions. Meta's own average invalid traffic runs 8.20% per Fraudlogix 2026. Instagram: 38%. Audience Network: 67%. A bot that clicked your ad, landed on your page, and completed your purchase flow produced a conversion event. It reached your attribution window. It became a training example for Andromeda. The algorithm now has a partial profile of that bot's traffic pattern and is looking for audiences that resemble it.
Stack these. Andromeda is partly blind to your real buyers and partly trained on fake ones. It does its job with total confidence. It optimizes. It finds more traffic that looks like its training set. Some of that training set is bots. Budget flows toward the ghost audience, systematically, while every attribution setting in your dashboard reads correct.
This is why accounts running clean attribution settings still see CPAs climbing, lookalike audiences drifting, and Advantage+ spending on traffic that does not convert to revenue. The settings were never the problem. The events inside them were.
What each window setting actually does
7-day click / 1-day view (default)
Captures the realistic consideration window for most ecommerce. A customer sees the ad on Monday, thinks about it, clicks on Wednesday, buys on Thursday: credited. A customer sees the ad, does not click, buys the same day from a different session: credited via view-through. This is the right starting point for most accounts. The view-through component requires clean traffic: Instagram at 38% IVT means a meaningful share of view-through credits are bot impressions.
7-day click / no view-through
Removes the view-through signal entirely. Cleaner for high-IVT placements. Less data volume, which can extend the learning phase for low-volume accounts. The right choice for B2B lead generation where view-through attribution inflates MQL counts and degrades lead quality signals back to the algorithm.
1-day click / no view-through
The strictest window. Cleanest signal per event. Lowest volume. Only viable for accounts with high daily conversion volume (50+ per day minimum for the algorithm to learn). For most SMBs, this starves the algorithm and produces unstable delivery.
28-day click (historical)
No longer available as a default in Ads Manager but accessible via the Attribution API for offline conversion uploads. Useful for long B2B sales cycles where the purchase decision span genuinely runs weeks. Contamination risk is proportional to window length: more days means more bot sessions that can fall inside the window relative to the conversion event.
View-through only
Avoid for any placement with meaningful IVT. Specifically avoid for Instagram and Audience Network. The view-through signal is the weakest causal connection between ad impression and conversion, and the easiest for bot impressions to contaminate.
CAPI solves the wrong half of the problem alone
The standard advice in 2026 is correct as far as it goes: set up the Meta Conversions API. CAPI moves conversion events server-side so they survive the blockers and ITP restrictions that drop 25-35% of pixel events. More real conversions reach the window. That is real and valuable.
CAPI is a more reliable pipe. It is not a filter.
When you send conversions server-side, the bot conversions travel the same pipe as the real ones. They arrive looking cleaner than browser-fired events, because server-side events carry less of the browser fingerprint that might have exposed them. You recover your real converters and deliver your fake ones with higher EMQ. The training set gets more complete and more contaminated at the same time.
A B2B SaaS company running a signup funnel analysis found 4,560 signups over four weeks. Inspected them. Only 730 were real humans. 84% fraudulent. 650 accounts from one laptop. Run that through Meta CAPI with a 7-day click window. Every fraudulent signup that arrived inside the window became a training example for Andromeda. The algorithm profiled the traffic patterns behind 3,830 bot sessions and started finding audiences that look like those patterns. The CAPI pipe worked perfectly. The water was poisoned before it entered.
The complete fix has three parts, in order.
First: collect conversion events from your own subdomain, not a third-party script. DataCops' first-party analytics loads from datacops.yourdomain.com. Not on any ad-blocker filter list. The events that were invisible to your pixel because uBlock blocked the script are now visible. First-party cookie lifetime extends from 7 days ITP to 90-400 days. The real buyers who were missing from your attribution window are present.
Second: filter non-human traffic before events fire. DataCops' fraud traffic validation runs IP intelligence against 361B+ network ranges, browser fingerprinting across 50+ signals detecting Puppeteer, Selenium, and Playwright headless automation, and email intelligence at the form layer. A bot session that passed every standard IP blocklist check is stopped at the server layer before its conversion event reaches the CAPI pipeline.
Third: enforce consent before events fire. Events from users who rejected consent should not carry identifiable parameters to Meta. DataCops' first-party CMP records consent on every session including the 30-40% where a third-party CMP would have been blocked, and propagates the consent state to the CAPI pipeline on the same server-side infrastructure.
Clean events reach your attribution window. Andromeda trains on real buyer patterns. The algorithm does its job and finds more people who look like your actual customers.
The EMQ connection
Event Match Quality measures how precisely Meta can match your conversion event to a Facebook user profile. EMQ 6.0 is considered healthy. Full enrichment with hashed email, phone, external_id, fbc, fbp, IP address, and user agent reaches 8.5-9.3 on Purchase events.
Going from EMQ 8.6 to 9.3 produces 18% lower CPA and 22% ROAS lift per Meta and TrackBee published data.
High EMQ on bot events is worse than low EMQ on real events. A bot event with EMQ 9.0 matches precisely to a Facebook profile. Andromeda records that profile as a high-confidence conversion example and targets similar profiles aggressively. The precision of the contamination is now working against you.
EMQ optimization and bot filtering are the same problem from different angles. You want high EMQ on clean events. You want zero events from bot sessions, high or low EMQ.
DataCops Business at $49/month handles both: bot filtering before events fire, then full enrichment of the surviving clean events via Meta CAPI, Google Ads Enhanced Conversions, TikTok Events API, and LinkedIn Insight CAPI from one pipeline.
Attribution window settings by account type
Ecommerce, under $50K/month GMV, primarily Meta:
7-day click / 1-day view. Standard window. Get CAPI live. Prioritize bot filtering over EMQ optimization if you have to choose. The volume loss from contaminated events hurts the algorithm more than slightly lower EMQ on clean ones.
Ecommerce, $50K-500K/month GMV, Meta and Google:
7-day click / 1-day view on Meta. Server-side CAPI on both platforms with unified bot filtering. DataCops Business at $49/month covers Meta CAPI, Google Enhanced Conversions, TikTok, and LinkedIn from one pipeline with 361B+ IP filtering before events dispatch.
Ecommerce above $500K GMV, Shopify-native:
Elevar at $200-950/month for deep Shopify Checkout Extensibility hooks and Shop Pay ClickID recovery. Elevar's order-level fidelity at this revenue level is worth the premium. DataCops for multi-platform CAPI and bot filtering if Elevar's Shopify-only scope is a constraint.
B2B lead generation:
7-day click / no view-through. View-through in B2B inflates MQL counts and degrades quality signals. HubSpot AI lead scoring for the offline conversion loop: feed pipeline-qualified leads back to Meta as offline conversions so Andromeda trains on real buyers, not raw form fills. SignUp Cops at the form layer to catch bot signups before they enter the CRM and the CAPI feed.
High-IVT verticals (finance, legal, home services):
Disable Audience Network for conversion campaigns. Audience Network IVT runs 67%. Instagram IVT runs 38%. Limit delivery to Facebook Feed, Facebook Stories, and Reels only. Use 1-day click / no view-through to minimize the surface area for contamination. Bot filtering is mandatory, not optional.
When DataCops is not the attribution answer
For Shopify stores above $500K GMV where millisecond purchase event accuracy and Shop Pay ClickID recovery matter more than multi-platform CAPI: Elevar's native Shopify integration is built for that specific problem. DataCops is a universal first-party pipeline. Elevar is Shopify-native infrastructure with checkout-level fidelity.
For enterprises with in-house GTM engineers who want full container control: Stape at $17/month Pro for sGTM hosting gives complete flexibility. DataCops is an outcome. Stape is infrastructure. Both are valid; they serve different buyer profiles.
For pure attribution dashboarding across channels: Triple Whale at $179/month annual, Northbeam at $1,500/month entry, or Rockerbox for multi-channel unified attribution. DataCops cleans the events before they reach those platforms. It does not replace the dashboards themselves.
For SOC 2 Type II certification required today from every vendor in your stack: DataCops is completing it. Tracklution holds SOC 2 and ISO 27001 active.
The Andromeda timeline
Before October 2025: contaminated attribution signals degraded campaign performance gradually. Accounts could run weeks of bot-contaminated conversions before the targeting drift became detectable in ROAS. Experienced media buyers had weeks to diagnose and correct.
October 2025: Project Andromeda fully deployed. The algorithm that acts on your attribution signals now acts within hours.
The contaminated conversion you sent to Meta CAPI yesterday was studied by Andromeda this morning. The targeting profile it contributed to is active now. If that conversion came from a bot session, Andromeda is already looking for more traffic that resembles the bot's patterns.
Your attribution window settings have not changed. The speed of the consequence has.
Your current 7-day click window contains every conversion that fired inside it over the past week. Andromeda has studied all of them and built a targeting profile from what it found.
How many of those conversions came from sessions that were real humans, with valid consent, whose traffic patterns you actually want Andromeda optimizing toward?
