The learning phase is the period after an automated campaign launches or is materially changed, during which the platform's bidding and delivery system collects enough outcome data to predict which impressions are worth buying. Until that data arrives, spend is spread across audiences, placements and bid levels the system cannot yet price with confidence, and costs swing. Google Ads and Meta Ads Manager both display the state as a status label, "Learning", and remove it once predictions settle.

The phase exists because conversion-optimised bidding runs on probability estimates. A new campaign, a new conversion goal or a changed audience weakens those estimates, and the label is how platforms acknowledge the gap. It is also a warning: early results do not represent steady-state performance.

How the calibration works

Every automated bid combines a prediction with a value. A system optimising for purchases estimates, auction by auction, how likely a given user in a given context is to buy, then bids according to the advertiser's target. At launch those estimates rest on priors such as account history, similar advertisers and contextual signals. The system deliberately buys some impressions where it is uncertain, a trade-off known in machine learning as exploration versus exploitation. Each recorded conversion narrows the uncertainty until the platform shifts toward exploiting what it has learned.

Meta applies the rule at ad set level. According to Meta's Business Help Centre, an ad set generally needs about 50 optimisation events within seven days of its last significant edit to exit learning. The optimisation event is whatever the advertiser chose to optimise for, so a purchase-optimised ad set needs purchases rather than clicks. Ad sets that fall short are labelled Learning Limited. Changes to targeting, placements, the optimisation event or bid strategy count as significant edits, as do adding a new ad, pausing for more than seven days and large budget or bid changes. The Marketing API exposes the state through a learning_stage_info field on each ad set.

Google attaches the status to bid strategies instead. Google Ads shows "Learning" after a new or reactivated strategy, a setting change, a composition change such as adding or removing campaigns, ad groups or keywords, or, in Shopping, an ad group target change. According to Google's help centre, calibration "can take up to around 50 conversion events or 3 conversion cycles", a conversion cycle being the usual delay between click and conversion. Manual CPC has no learning period. Google also says its algorithms keep learning after the label disappears and that a new cost per acquisition (CPA) or return on ad spend (ROAS) target does not discard what the model already knows.

Elsewhere the numbers vary. TikTok's "About Learning Phase" page says volatility typically declines after around 25 campaign results or seven days, while its FAQ still calls 50 conversions "the most significant indicator". Pinterest says its Learning mode lasts two weeks on average. Microsoft Advertising says portfolio bid strategies need "a few weeks". Reddit recommends 14 days without changes after creation or major edits for its Max campaigns. In Display & Video 360 (DV360), Google's demand-side platform (DSP), automated bidding "can take up to 4 weeks to learn and calibrate", according to its help centre, while custom bidding models need at least 10,000 scored impressions and 500 positively scored impressions per advertiser before a training run of one to three days.

Origin and evolution

The split between learning and optimised spend has programmatic roots. Documentation for Xandr, the former AppNexus platform, describes each inventory node moving from a "learn" phase to an "optimized" phase once it records enough success events, a threshold buyers could set between one and 15, with a separate learn budget capping exploratory spend. Microsoft completed its purchase of Xandr in June 2022 and scheduled the Invest DSP's closure for February 28, 2026.

Google formalised the status in search. When it grouped its conversion-based strategies under the Smart Bidding name in 2016, AdWords added bid strategy statuses including "learning", according to a PPC Hero report from August 2016. A Google whitepaper from January 2018 stated that "it typically takes 1-2 weeks" to calibrate a new strategy and tied stability to volume: a Target CPA strategy with 30 conversions in 30 days learned slowly, one with 500 very fast, and Target ROAS needed at least 50.

Facebook gave the concept its name. In September 2017 Ads Manager began showing "Active (learning)" on ad sets. "The learning phase is a normal part of the lifecycle of an ad set. It's always existed, we're now just letting you know it's happening," the company said, according to Social Fire Media, which reported that the recommended volume had risen to 50 purchases from the 15 to 25 previously advised. Learning Limited followed in 2019, flagging ad sets unlikely ever to gather enough data.

The direction since has been toward lower thresholds and pooled data. Meta advertising tool vendors reported in June 2024 that the requirement had fallen to 10 events for purchase and app install campaigns, and in August 2026 that Advantage+ shopping campaigns could exit after 25 purchases. Neither report was accompanied by public Meta documentation. Google product managers said in March 2026 that advertisers no longer need a bank of conversion data before starting Smart Bidding, because the system trains on every conversion in an account.

Why it shapes account structure

Spend during learning buys less efficiently, and every reset repeats that cost, which has changed how accounts are built. On Meta, fewer ad sets with larger budgets reach 50 events sooner; one advertiser lifted ROAS from 1.8 to 3.2 by merging three $50-a-day ad sets into a single $150-a-day Advantage+ campaign. At Google, Brandon Ervin, director of product management for search ads, said in February 2026 that merging ad groups or campaigns with similar creative should not require much of a learning period, because models key on semantic features rather than campaign IDs. Switching conversion actions or adopting AI Max, he said, does.

Platform guidance now writes freeze windows into launch plans. Google asks Demand Gen advertisers on value-based bidding to make no bid or creative changes for 14 days, after qualifying with 50 conversions with value in 35 days, and gives App campaigns the same 14-day instruction. AI Max for Search carries a learning period of one to two weeks and a $50 daily budget floor. Microsoft Advertising says search themes shorten the Performance Max learning period.

Evaluation windows follow. Google advises judging Target CPA over at least 30 conversions and Target ROAS over 50. Ginny Marvin, Google's Ads Product Liaison, advised in 2024 assessing lead-generation bidding over 50 conversions or a full month, excluding ramp-up.

Automation has added a new source of resets. When Meta opened AI Connectors for Claude and ChatGPT in April 2026, Brody K., a client partner at Meta, wrote on LinkedIn: "Just be sure to remind your agent about the learning phase."Software with write access can trigger repeated resets through frequent edits.

Limitations and disputes

The thresholds are heuristics set by the seller. Facebook tripled its benchmark in 2017 without publishing the evidence, and later reductions have surfaced through vendors rather than documentation. TikTok's own pages disagree on whether 25 results or 50 conversions matter.

Low-volume advertisers carry the heaviest cost. Business-to-business marketers, high-ticket retailers and lead generators may never reach 50 weekly purchases. The standard remedy, optimising toward an earlier event, changes what the system buys: an ad set optimising for add-to-cart learns to find people who add to carts. Ervin and Marvin both suggested earlier-funnel actions for campaigns below roughly 15 conversions in 30 days.

Incentives are awkwardly aligned. Larger budgets, fewer ad sets, broader targeting and fewer manual controls are the fixes platforms recommend, and each also suits their commercial interest while reducing advertiser visibility. The label can become a reason to keep spending through weak results; Google itself stresses that learning status does not mean a campaign is failing.

The label and the learning also diverge. Conversion lag of seven to 21 days in complex sales means a status can clear before the data it depends on is reported. Google's August 2026 change to budget-limited campaigns exposed the same gap: target changes take one to two conversion cycles, often around seven days, to settle. No independent body measures what learning periods cost advertisers.

Not the same as

Cold start is the general machine learning problem of predicting without data. The learning phase is a platform-declared, time-bounded status built around it.

A/B test results come from a controlled split with a holdout. The learning phase has no control group, and performance within it is not comparable to steady state.

Model training, as in the "Training" status of a DV360 custom bidding model, happens before serving; assigning an untrained model halts spend. Learning happens with live money.

Limited, as a Google bid strategy status, means constraint by budget, bid limits or inventory rather than insufficient data, and is distinct from Meta's Learning Limited.

Recent developments

After relabelling Target CPA and Target ROAS in June 2026 without touching bidding logic, Google altered behaviour for budget-limited target campaigns from August 17. Marvin said that update would not by itself change spend. Smart Bidding Exploration, expanding to Performance Max and Shopping, turns exploration into an explicit setting with a ROAS tolerance of 5% to 30%. Demand Gen results released in August 2026 still assume at least 50 conversions during the learning period. On September 22, 2026, Search Engine Land reported that Google had reframed its learning period guidance for Search, Shopping and Performance Max around conversion volume rather than elapsed time.

Timeline

  • August 2016: Google groups conversion-based strategies under Smart Bidding and adds bid strategy statuses, including "learning"
  • September 2017: Facebook Ads Manager begins showing "Active (learning)" and raises its recommended volume to 50 purchases
  • January 2018: A Google whitepaper says new bid strategies typically take one to two weeks to calibrate
  • 2019: Facebook introduces the Learning Limited delivery status
  • June 2022: Microsoft completes its acquisition of Xandr, whose optimisation engine separates learn and optimised spend
  • June 2024: Meta tool vendors report the learning threshold falling to 10 events for purchase and app install campaigns
  • August 2024: Ginny Marvin advises evaluating value-based bidding for lead generation over 50 conversions or a month
  • June 2025: Google sets 50 conversions with value in 35 days as the Demand Gen value-based bidding threshold
  • October 2025: AI Max for Search guidance sets a one-to-two-week learning period and a $50 daily budget minimum
  • January 2026: Reddit recommends 14 days of stability for Max campaigns after launch or major edits
  • February 2026: Brandon Ervin says consolidating similar campaigns should not require much of a learning period
  • February 28, 2026: Scheduled closure date for Microsoft Invest DSP
  • March 2026: Google product managers say Smart Bidding no longer needs a bank of conversion data to start
  • April 29, 2026: Meta opens AI Connectors, with a warning that agents can trigger repeated learning phases
  • June 16, 2026: Google announces the relabelling of Target CPA and Target ROAS, with no change to bidding logic
  • August 2026: Vendors report Advantage+ shopping campaigns exiting learning after 25 purchases
  • August 17, 2026: Google changes delivery for budget-limited campaigns using CPA or ROAS targets
  • September 22, 2026: Search Engine Land reports Google reframing its learning period guidance around conversion volume

Summary

Who. Platforms define and display the learning phase: Meta at ad set level, Google Ads, Microsoft Advertising and DV360 at bid strategy or line item level, with TikTok, Pinterest and Reddit publishing their own rules. Advertisers, agencies and, increasingly, AI agents trigger it through launches and edits, and pay for the less efficient spend it involves.

What. A calibration period during which an automated bidding or delivery system gathers conversion data to predict which impressions are worth buying. Performance fluctuates until the platform judges its estimates stable, typically after about 50 conversions or a few conversion cycles, depending on the platform.

When. Programmatic optimisation engines separated learn and optimised spend before social platforms named the concept. Google added a learning status in 2016, Facebook introduced the learning phase label in September 2017 and Learning Limited in 2019. Thresholds and guidance have kept shifting through 2024, 2025 and 2026.

Where. Inside the bidding systems of search, social and programmatic platforms: Google Ads, Meta Ads Manager, TikTok Ads Manager, Pinterest, Reddit, Microsoft Advertising and DV360, and in the APIs and agent connectors that now edit campaigns on advertisers' behalf.

Why. The learning phase determines how campaigns are structured, when they can be changed and how long before results can be judged. Its thresholds are set by the platforms selling the media, published inconsistently and rarely audited, which makes it one of the least transparent costs in automated buying.