@akshitraja

What they said

1. Meta serves an ad in two steps.
2. Step 1 = retrieval engine called Andromeda. It shortlists "a few thousand" ads for that user out of "crores" (tens of millions) of ads.
3. Step 2 = auction/ranking among the few thousand remaining. The winner is picked depending on bid, campaign budget and ad quality (caption also says estimated action rates).
4. If your ad is not selected in step 1, bid and budget do not matter at all ("फरक ही नहीं पड़ता").
5. How shortlisting works:
   • Andromeda converts your creative into a vector. AI models understand numbers, not words or images; a vector is an array of numbers holding the semantic meaning: your proposition, pain point, messaging, visuals (the animation labels the four: Proposition, Pain point, Messaging, Visuals; example vector values like 0.82, 0.13, 0.67).
   • The creative vector is plotted as a point on an "embedding map".
   • The user is also converted to a vector (interests, behaviour, demographics, per the diagram labels) and plotted on the same map. (Transcript 30-60s is partly truncated in the file; this part is reconstructed from the frames and caption.)
   • More similarity = less distance between points. That is how targeting/shortlisting is defined.
   • Worked example: user is point E; creatives that sit very near (point A and neighbours) closely match their interests and demographics, so those get shortlisted. Far-away creatives are "ignored" (label on frame 76-78s).
6. Closing: "this is the first step, called Andromeda, the retrieval mechanism to shortlist the ads."
7. Caption adds: genuinely different creative concepts matter more than ten minor variations of the same ad. Teases a next video on "what makes two creatives meaningfully different for Meta".

Try this

• Meta advertisers: Build 4 to 6 creatives that differ in concept, not in colour or crop. Change the customer motive (gift, self-reward, bundle, occasion), the pain point or the format (UGC, product-only, founder story).
• Sellers with several products or benefits: Give each pain point or benefit its own ad. Each one lands in a different part of the audience, so you reach more people than ten edits of one hook would.
• Test properly: Judge each concept on cost per purchase after enough spend, and retire concepts that look alike in the first seconds.
• Agencies and freelancers: Use the two-step model (retrieval, then auction) to explain to clients why you want concept variety in the creative budget, not just more edits.
• Creators: A talking head over a simple animated diagram is a cheap, repeatable explainer format. End with a teaser for part two.
• Keep perspective: It is a simplified model, so check Meta's engineering blog before acting on the details.

Links

• Meta Andromeda (Meta's ad retrieval engine). Source link given by creator in a comment reply: https://engineering.fb.com/2024/12/02/production-engineering/meta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine/ (Meta Engineering blog; creator says it says retrieval narrows tens of millions of candidates to a few thousand before ranking).
• Hashtags: #MetaAndromeda #MetaAds #FacebookAds #EcommerceMarketing #PerformanceMarketing.
• @trialandai tagged as co-author.
• Music: "snowfall" by Øneheart, reidenshi (background).
• A commenter suggested salfars.com (some tool; irrelevant, creator replied "No i will try phot.ai").