Guide · AI for your ads · For the owner whose WhatsApp chats start with an ad
Why Meta shows your WhatsApp ad to one person and not another: the ad auction, Meta's ranking AI, the learning phase and the edits that reset it
If your WhatsApp chats start with an ad, before your reply speed or your prices can matter, Meta has to decide whether to show the ad at all, and to whom. Meta says that decision is made by an auction and by machine learning, its AI that predicts what each person will do. This guide puts Meta's Business Help Center and its own March 2025 announcement about its ad ranking AI into plain words: what wins the auction, what the AI is predicting, why a new ad goes through a learning phase, which edits start that learning again, and the Ads Manager labels that tell you how it is going.
Published 22 October 2026
The short version
Meta's help article on ad delivery opens with one sentence that explains most of what follows: “The Meta ad delivery system uses an ad auction and machine learning to determine where, when and to whom we show your ads.”
- The auction
- Every time there is a chance to show someone an ad, Meta runs an auction to decide which ad that person sees.
- The machine learning
- Meta's systems predict how likely that person is to find each ad worthwhile and to do what the advertiser wants (Our reading: for your ad, start a chat). Those predictions count in the auction alongside the money.
- The learning phase
- A new or significantly edited ad set starts with little for the system to go on. While it gathers enough results, Meta marks it as Learning, and its results swing more than usual.
What wins the auction
Meta's article on ad auctions says the ads that can compete for a person are the ones whose target audience includes that person. The winner is the ad with the highest total value, subject to a price floor (a minimum price) that may affect whether an ad is shown and the price paid. Total value combines three things:
- Bid
- What the advertiser is willing to pay to get the outcome they asked for. Meta offers several bid strategies for managing it.
- Estimated action rates
- Meta's estimate of the probability that showing this ad to this person leads to the outcome the advertiser wants. Meta adds that clickbait and engagement bait do not improve ad performance.
- Ad quality
- Measured from many sources, including people's feedback when they view or hide the ad, and checks for low-quality signs such as withholding information, sensationalised language and engagement bait.
The article then says: “Together, estimated action rates and ad quality measure ad relevance.” Because both count, “an ad that's more relevant to a person could win an auction against ads with higher bids.” It also says Meta may use data about you and your ad account to adjust auction outcomes and prices, for reasons it lists such as ad quality, testing ad products and complying with the law, but that “such auction adjustments will not cause us to charge you more than your bid”.
Our reading
Where Meta's AI comes in
The auction article talks about estimates; Meta's announcement of 27 March 2025, AI Innovations in Meta's Ad Ranking Driving Advertiser Performance, described the AI systems behind its ad ranking. It named four:
- Meta Andromeda
- Introduced in 2024. Meta said it powers the more complex models used for ads retrieval, which it called the first step in the ranking process: narrowing the full pool of ads to the shortlist Meta considers showing someone. Meta said that as businesses upload more creatives, Andromeda helps it pick the right creative for each person.
- GEM, the Generative Ads Recommendation Model
- A machine learning model that Meta said lets its ads system rapidly process large amounts of data to deliver relevant, personalised ads. Meta said it had rolled GEM out more broadly in its ads system after results in Meta Reels.
- Meta Lattice
- Which Meta called its ad ranking architecture: fewer, larger models that learn across campaign objectives and surfaces, in place of many smaller models each tuned for one objective or surface.
- Sequence Learning
- Looks at the order of things a person does before and after seeing an ad. Meta's example: after someone books a ski resort room, show ski equipment or lift tickets rather than more ski resorts.
The delivery article adds how this behaves day to day. Each time an ad is shown, the system's predictions become more accurate. If you have several ads, Meta shows the one most likely to get the result at the lowest cost for that particular person, so your ads will not necessarily be shown the same number of times. And it decides where to deliver next from “predictions of future performance”, not from each ad set's past results.
Our reading
The learning phase
Meta's article on the learning phase describes it as the period when the delivery system still needs to learn how an ad set may deliver and perform, after you create a new ad or ad set or make a significant edit. The Delivery column in Ads Manager reads Learning while it lasts. An ad set leaves the learning phase as soon as it can deliver stably, which the article says usually happens after about 50 results in the week after the ad set's last significant edit.
To check progress, it says to open the ad set in Ads Manager and look at the Last significant edit column and the Results column, adding them through Columns if you do not see them. During learning, results are less stable and cost per result is usually higher. Meta's advice:
- Wait to edit
- Editing during the learning phase resets learning and delays the system's ability to optimise.
- Avoid unnecessary edits
- Only edit when you have reason to believe the edit should improve performance.
- Avoid high ad volumes
- With many ads and ad sets the system learns less about each one; combining similar ad sets combines learnings.
- Use realistic budgets
- A very small or inflated budget misleads the system, and frequent budget changes can restart learning.
The article ends: “you shouldn't try to avoid the learning phase completely.” Testing new creative is how performance improves over time.
Where Meta's pages differ
Which edits start the learning again
Meta's article on significant edits says every edit has some effect on delivery, but only a significant edit sends an ad set back into the learning phase. It lists these as significant:
- Targeting
- Any change to who the ad set is aimed at.
- Creative
- Any change to the ad itself, such as the picture, video or text.
- Optimisation event
- Any change to the result the ad set is told to get.
- A new ad
- Adding a new ad to the ad set.
- A long pause
- Pausing the ad set for 7 days or longer; it goes back into learning when you switch it on again.
- Bid strategy
- Changing the bid strategy.
Changes to the budget, the spending limit, or a bid, cost per result or return on ad spend goal may or may not count, depending on how big the change is: a small change is unlikely to restart learning, a large one may. With Advantage+ campaign budget, the article says Meta moving budget between ad sets does not restart learning, a significant edit made to one ad set does not send the campaign's other ad sets back, and nor does adding a new ad set; changing the campaign's bid strategy or budget might send several ad sets back at once.
Our reading
Learning limited is not a penalty
If an ad set is unlikely to get about 50 optimisation events in the week after its last significant edit, the Delivery column reads Learning limited. Meta's article says plainly: “Learning limited isn't a penalty”. It means the system cannot optimise well with the current setup. The causes it lists are a small audience, a low budget, a low bid or cost control, high auction overlap (Our reading: your own ad sets competing for the same people), a result that happens rarely, or too many ads running at once.
Its fixes: combine ad sets and campaigns, expand the audience, raise the budget, raise the bid or cost control, or choose a result that happens more often (Meta's example is moving from purchases to add to cart). You can add Optimization events as a column to see how close you are.
Our reading
The three relevance rankings in Ads Manager
Meta's article on ad relevance diagnostics says they replaced the old ad relevance score and compare each ad's past performance with ads competing for the same audience:
- Quality ranking
- How your ad's perceived quality compared with those ads.
- Engagement rate ranking
- How your ad's expected engagement rate compared.
- Conversion rate ranking
- How your ad's expected conversion rate compared with ads that have the same optimisation goal.
They show by default in the Performance column preset. Otherwise, in Ads Manager click the Ads tab, Columns, Customize columns, select the three rankings and click Apply. They are not shown for ads with fewer than 500 impressions, and the article adds: “Ad relevance diagnostics aren't inputs into the ad auction.”
Our reading
When people have seen the same picture too often
Meta's article on creative fatigue says it happens when an audience has seen the same creative too many times and becomes less likely to engage, which can push up cost per result. Meta's recommendations for it are available for ad sets with one creative, with some exclusions the article lists. Before launch, Meta may warn you if it predicts fatigue in the first 7 days. After launch, the Delivery column can show Creative limited, when cost per result is higher than your past ads but less than twice as much, or Creative fatigue, when it is twice as much or more. Meta counts recent showings of the same image or video, including in other campaigns from your Page.
Its recommendations: create another ad with an image or video that is materially different (keeping the original running may maximise results), expand the audience, or try Advantage+ creative, which the article says is only available for campaigns with the traffic or sales objective and a website destination. The same article says opportunity score, where these recommendations can appear, does not reflect your actual or future performance.
Our reading
What these pages do not say about WhatsApp
None of the eight Meta pages behind this guide names WhatsApp, ads that open a WhatsApp chat, the WhatsApp Business app, the UAE or any Gulf country. They describe Ads Manager. They do not say how the learning phase or the rankings appear for an ad made from inside the WhatsApp Business app.
Our reading
What this means for your business
BitByteCore is a verified Meta Tech Provider (since 9 September 2026). This guide repeats Meta's pages as we read them on 22 October 2026, and is not legal advice.
Our reading
Questions
Does the highest bid always win Meta's ad auction?
No. Meta's Business Help Center says the winner is the ad with the highest total value, which combines the bid, Meta's estimate of whether the person will take the action the advertiser wants, and ad quality. It says an ad that is more relevant to a person could win against ads with higher bids, and that auction adjustments will not charge you more than your bid.
Is Meta's ad delivery really decided by AI?
Meta's help article on ad delivery says the delivery system uses an ad auction and machine learning to decide where, when and to whom ads are shown. Meta's announcement of 27 March 2025 described four AI systems in its ad ranking: Meta Andromeda, GEM, Meta Lattice and Sequence Learning.
What does Learning mean in the Delivery column?
That the ad set is in the learning phase: the system is still working out how best to deliver it, after you created it or made a significant edit. Meta says ad sets usually leave it after about 50 results in the week after the last significant edit, and that results are less stable and cost per result usually higher meanwhile.
Which changes send my ad back to learning?
Meta lists any change to targeting, creative or the optimisation event, adding a new ad, pausing the ad set for 7 days or longer, and changing the bid strategy. Budget and bid amount changes may or may not count, depending on how big the change is.
Is Learning limited a penalty?
No. Meta says it is not a penalty but a sign the system cannot optimise with the current setup, for reasons such as a small audience, a low budget, a low bid or cost control, high auction overlap (Our reading: your own ad sets competing for the same people), a rare result or too many ads. Its fixes include combining ad sets, expanding the audience and raising the budget.
Do the quality, engagement and conversion rankings decide who sees my ad?
No. Meta's article on ad relevance diagnostics says they are not inputs into the ad auction. They compare your ad's past performance with ads competing for the same audience, and are not shown for ads with fewer than 500 impressions.
Who is BitByteCore?
BitByteCore is a verified Meta Tech Provider (since 9 September 2026). This guide repeats Meta's own pages, and is not legal advice.
Sources
- Meta Business Help Center: About ad delivery, read 22 October 2026.
- Meta Business Help Center: About ad auctions, read 22 October 2026.
- Meta for Business: AI Innovations in Meta's Ad Ranking Driving Advertiser Performance (27 March 2025), read 22 October 2026.
- Meta Business Help Center: About the learning phase, read 22 October 2026.
- Meta Business Help Center: Significant edits and learning phase, read 22 October 2026.
- Meta Business Help Center: About learning limited, read 22 October 2026.
- Meta Business Help Center: About ad relevance diagnostics, read 22 October 2026.
- Meta Business Help Center: About creative fatigue recommendations in Meta Ads Manager, read 22 October 2026.
This guide describes Meta's pages as we read them, and is not legal advice. More plain-language notes like this one are on the guides page.
Want customer conversations in one place you own?
We set up a business's WhatsApp number on a system its team runs, so every customer conversation sits in one place the business owns. A short call is the easiest way to see how that would work for you.
Book a demo