Meta Ads Creative: Does Meta Read the Text on Your Ads?

Meta ads creative explained: what Meta has confirmed about reading text in ad images, which claims are myths, and how to write on-image copy that works.

Updated 11 min readBy the Static Ad Templates team

Key takeaways

  • On Oct 19, 2026, Meta said its pre-trained multimodal models read each ad, including its text and images, and encode it into an embedding used for retrieval and ranking.
  • Meta has not said Andromeda runs OCR on ad images, so the claim that the words on your static shape delivery is a strong inference, not a Meta statement.
  • Meta's own creative guidance warns that ads that are too similar may not be recognized as different and may be delivered to the same audiences.
  • Write on-image headlines that name one persona, pain or benefit, and change the visual and format along with the words for each new ad.
  • Entity IDs, a 60% similarity threshold and the 20% text rule are not current Meta rules; the 20% rule was retired in 2020.

Probably, yes. Meta has never said Andromeda runs OCR on ad images. But on October 19, 2026, Meta confirmed its multimodal models "read each ad, product, and page, including its text and images," and that similar items get grouped together. So the words on your static very likely shape who sees it. Write them on purpose.

The short answer, in one table

Most posts on this topic either overclaim ("Andromeda reads your headline and targets people") or dismiss it. Here is where the evidence actually sits as of November 2026.

Claim Status Basis
Meta's ad models read ad text and images to represent each ad Confirmed Meta Engineering, Oct 19, 2026
Meta uses LLMs to reason about ad content during ad retrieval Confirmed Meta Q2 2026 earnings call
Too-similar ads may be treated as one and shown to the same people Confirmed Meta's creative differentiation guide
Andromeda specifically runs OCR on your image Not stated by Meta No Meta source says this
The words printed on your static influence who sees it Strong inference Follows from the confirmed items above
On-image text works like keyword or interest targeting No It is one input to a prediction

If you want the background on the retrieval system itself, start with our Meta Andromeda explainer. This post zooms in on one question about Meta ads creative: how much of your ad's content Meta's AI understands, and what that means for the copy you put on your images.

What Meta has actually said, in order

The evidence is a chain of dated statements. No single one says "we read the headline on your image," but together they get close.

December 2024: Andromeda is a retrieval engine

Meta announced Andromeda on December 2, 2024, on its engineering blog. It is the retrieval step: it narrows "tens of millions" of eligible ads down to "a few thousand" candidates, and later ranking models plus the auction pick what actually shows.

That post mentions "precomputed ad embeddings" but does not describe analyzing the image or text of a creative. It does contain the line that started the diversification push: "Increased ad diversity can improve people's experience with ads and drive better advertiser outcomes."

The October 28, 2026 follow-up post (the one the industry calls "Andromeda 2.0," a label Meta does not use) lists "LLM-derived semantic features" among its inputs and says ad content is fused with serving context. It still does not mention OCR or visual analysis.

November 2025: GEM uses a "creative representation"

GEM is Meta's ranking-stage foundation model. Meta's November 10, 2025 engineering post lists "creative representation" among its features, alongside age, location and ad format, and says GEM learns from interactions "across modalities such as text, images, audio, and video."

Meta reported that GEM drove a 5% increase in ad conversions on Instagram and 3% on Facebook Feed in Q2 2025. Those are Meta's own numbers.

July 2026: LLMs reason about ad content inside retrieval

On Meta's Q2 2026 earnings call, CFO Susan Li described the Meta Generative Recommender as "a paradigm shift in how our ads system works. Rather than scoring every possible ad individually, we are now using LLMs to reason about ad content and user preferences together, and predict the best ad for each person." She added: "We deployed the first generative model into our ads retrieval system."

Retrieval is the stage Andromeda handles. So by mid-2026, content understanding sits in the same part of the system that decides which ads are even considered for a person.

October 19, 2026: "including its text and images"

This is the most direct statement Meta has made. In its GenRecs Tokenization post, Meta wrote: "Pre-trained multimodal models read each ad, product, and page, including its text and images, and encode the items into an embedding that carries world knowledge."

Those embeddings are turned into tokens, and "Similar items share tokens, much as related words share roots." Meta says the system is used for "both retrieval and ranking" and reported a +0.35% lift in conversions across surfaces (Meta internal data, measured February 2026).

Meta's guidance: too similar means same audience

Meta's creative differentiation guide for small businesses puts the practical consequence plainly. The English version, as widely quoted by practitioners, reads: "When creative is too similar, our system may not recognize it as different and could deliver it to the same sets of audiences," and that overlap "limits your reach." We read the Swedish-locale version of the page, which says the same thing.

Background: Meta has read text in images since 2018

Meta has had the capability for years. Its Rosetta system, described in September 2018, extracts text from "more than a billion public images per day" for photo search, hate-speech detection and News Feed photo classification. That post does not mention ads, so treat it as proof of capability only.

What Meta has not said (does Facebook read text in images?)

No Meta source says Andromeda, or any ads system, runs OCR on your image as a separate step. The two Andromeda engineering posts do not describe reading creative content at all.

The inference is still strong. A multimodal model that reads an ad's "text and images" is looking at the image itself, and on a static ad the headline is part of the image. Modern image models of this kind generally pick up text rendered in a picture. So "For runners with knee pain" printed on your creative very likely ends up in how Meta represents that ad. That is our reading of the evidence, not a Meta quote.

"Targeting" is also the wrong word. Meta's system predicts which ad each person is most likely to act on. Your on-image headline does not select an audience. It changes what the ad "is" in Meta's representation, which changes who the system predicts it is relevant for.

Why "creative is the new targeting" follows

The exact phrase is practitioner shorthand, but Meta endorses the substance. In an April 22, 2025 post, Meta defined creative diversification as "creating a wide range of ad creatives with different themes, messages, and visuals to cater to diverse audience segments," and called it "the best lever to find the most relevant audiences."

Put that next to the GenRecs statement and the logic is simple. If Meta matches ads to people based on what the ads contain, then the message in each ad is the closest thing to an interest setting you still control. Four ads with four clear messages give the system four distinct things to match. Four ads with one message give it one. Our guide to creative diversification on Meta covers how many ads that means for your budget.

How to write on-image copy Meta's AI can use

These are practical rules that follow from the evidence above. None of them needs a hack or a hidden setting.

Name the persona, pain or benefit in plain words

"For runners with knee pain" tells a reader, and very likely Meta's models, exactly who the ad is for. "Move better every day" could be for anyone, so it gives the system nothing distinctive to match on.

Use concrete nouns: the activity, the situation, the outcome. "Squat day," "steep descents," "after a 10-hour shift" all carry meaning. Brand slogans usually don't.

One policy caution. Meta's personal attributes policy restricts ads that assert or imply something about the viewer, such as health conditions ("Are you depressed?" or "Your knee pain is getting worse"). Describe who the product is for in the third person, and avoid health claims you can't back up. Weight-loss before-and-after images carry extra rejection risk.

Change the message when you change the ad

A color swap, a new font or a different background is very likely the same ad to Meta. Meta's own definition of diversification names "themes, messages, and visuals," and Jon Loomer put it bluntly in October 2025: "Unless they're distinctly different enough, any variation is unlikely to make a meaningful impact on performance." He added that text differences "should be significant, not subtle."

Change the headline and the visual together. A new persona deserves a new scene, a new person in frame, and often a new layout.

Line up the image, primary text and landing page

If your image says "For hikers," the primary text should talk about trails and the landing page hero should show one. Consistent signals likely give Meta a cleaner read on who the ad is for. They certainly give the person who clicks a page that matches what they were promised.

Mixed signals blur the ad. An image aimed at runners with primary text about office workers and a generic homepage gives the system three different stories.

Skip tiny-variation spam

Uploading 30 versions with one word swapped feels like testing. Under a system where "similar items share tokens," those versions likely collapse into one concept and split its budget. Common Thread Collective made the same point in May 2026: "Run 10 variations of the same static product shot and they don't get 10 chances at distribution."

Four genuinely different concepts will teach you more than 20 lookalikes.

Worked example: one knee sleeve, four personas

Take a hypothetical compression knee sleeve. Here are four static ads built for four personas. Each one changes the words, the scene, the person and the format.

Persona On-image headline Visual Primary text opens with Landing page match Format
1 Runners "For runners with knee pain" Runner mid-stride on a road, sleeve visible "Built for the back half of your long run." Hero shows running; sizing guide by calf and thigh Problem + Solution
2 Hikers "Steep descents. Steady knees." Hiker on a downhill trail, sleeve over leggings "Downhill is where knees take the hit." Trail photo; breathability and fit under layers Feature/Benefit
3 Lifters "Squat-day support that stays put" Barbell back squat in a gym, side angle "No rolling down mid-set." Gym photo; silicone grip band close-up Us vs Them (vs. sleeves that slip)
4 Active 50+ "Gardening, stairs, grandkids" Person kneeling in a garden bed A real customer quote about daily use Everyday-wear section; easy on and off Testimonial

Now compare a weak set: the same product photo on four background colors, each reading "Premium Knee Sleeve. 20% Off." Based on Meta's own guidance, that set likely reads as one ad. It has one message, one visual idea and no persona.

Row 4 uses a testimonial, which brings its own rules. Use real customer words with permission. The FTC's final rule on fake reviews, announced in August 2024, bans fake testimonials, including AI-generated ones that pretend to come from real customers. Our testimonial ads guide covers how to source and lay these out.

If you're not a designer, layout is where this gets slow. In the Static Ad Templates library you can filter by format (Problem + Solution, Feature/Benefit, Us vs Them, Testimonial) so each persona gets a structurally different layout, then edit it in Canva. Our static ads guide covers the formats in more depth.

Myths to ignore

A lot of confident advice about Meta AI ads is vendor vocabulary dressed up as Meta policy. Here is how the common claims hold up.

Myth What the evidence says
"Andromeda runs OCR on your image." Meta hasn't said this. It has said its multimodal models read ad text and images (Oct 2026).
"Every ad gets an Entity ID, and similar ads share one auction ticket." "Entity ID" is an agency and vendor term. No Meta page uses it. The closest confirmed idea is that "similar items share tokens."
"Ads more than 60% similar get grouped together." No Meta source gives any similarity threshold. The 60% figure comes from vendor blogs.
"The 20% text rule still applies." Facebook removed the 20% text limit in September 2020, per Search Engine Journal. Its docs kept tips on reducing text as a best practice, but there is no cap.
"Andromeda raised ROAS by 22%." That figure appears in the December 2024 post but describes Advantage+, not Andromeda.
"Andromeda 2.0 is a new Meta product." It's an industry label for Meta's October 28, 2026 post, which says "next generation."
"On-image text is a targeting setting." It's one input to how the ad is represented. Delivery is still a prediction.

What to do next

  1. Audit your live ads. Write one line per ad describing who it's for and what it says. If three ads share the same line, you have one ad, not three.
  2. Pick three or four personas you can name in plain words, based on your reviews, support tickets or competitors' long-running ads.
  3. Write one explicit on-image headline per persona that names the person, the pain or the benefit.
  4. Build each ad with a different visual and layout, not just a different headline on the same photo.
  5. Match the primary text and landing page to each ad's persona.
  6. Launch them together in one broad ad set and judge results per ad over a full week. When a persona wins, write new angles for that persona instead of new colorways of the winner.

Sources

  1. Meta Engineering, "Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine" (Dec 2, 2024), https://engineering.fb.com/2024/12/02/production-engineering/meta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine/ (accessed Nov 2026)
  2. Meta Engineering, "Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design" (Oct 28, 2026), https://engineering.fb.com/2026/10/28/ml-applications/meta-andromeda-next-gen-ads-retrieval-gpu-co-design/ (accessed Nov 2026)
  3. Meta Engineering, "Meta's Generative Ads Model (GEM): The central brain accelerating ads recommendation AI innovation" (Nov 10, 2025), https://engineering.fb.com/2025/11/10/ml-applications/metas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation/ (accessed Nov 2026)
  4. Meta, "Q2 2026 Earnings Call Transcript" (Jul 2026), https://s21.q4cdn.com/399680738/files/doc_financials/2026/q2/META-Q2-2026-Earnings-Call-Transcript.pdf (accessed Nov 2026)
  5. Meta Engineering, "GenRecs Tokenization: Transforming Ads Recommendation at Meta with Stable Model Vocabulary" (Oct 19, 2026), https://engineering.fb.com/2026/10/19/ml-applications/genrecs-tokenization-transforming-ads-recommendation-at-meta-with-stable-model-vocabulary/ (accessed Nov 2026)
  6. Meta for Business, "Creative differentiation" (small business guide), https://www.facebook.com/business/m/small-business/creative-differentiation (accessed Nov 2026)
  7. Meta Engineering, "Rosetta: Understanding text in images and videos with machine learning" (Sep 11, 2018), https://engineering.fb.com/2018/09/11/ai-research/rosetta-understanding-text-in-images-and-videos-with-machine-learning/ (accessed Nov 2026)
  8. Meta for Business, "The Creative Advantage: unlocking the power of diversification with Meta Andromeda" (Apr 22, 2025), https://www.facebook.com/business/news/the-creative-advantage-unlocking-the-power-of-diversification-with-meta-andromeda (accessed Nov 2026)
  9. Jon Loomer, "Meta Andromeda and Creative Diversification: 7 Examples Explained" (Oct 20, 2025), https://www.jonloomer.com/meta-andromeda-creative-diversification/ (accessed Nov 2026)
  10. Common Thread Collective, "Meta Andromeda Killed Your ROAS. Here's the Only Way to Get It Back." (May 6, 2026), https://commonthreadco.com/blogs/coachs-corner/meta-andromeda-roas-creative-strategy-2026 (accessed Nov 2026)
  11. Search Engine Journal, "Facebook Removes the 20% Text Limit on Ad Images" (Sep 23, 2020), https://www.searchenginejournal.com/facebook-removes-the-20-text-limit-on-ad-images/381844/ (accessed Nov 2026)
  12. Federal Trade Commission, "Federal Trade Commission Announces Final Rule Banning Fake Reviews and Testimonials" (Aug 14, 2024), https://www.ftc.gov/news-events/news/press-releases/2024/08/federal-trade-commission-announces-final-rule-banning-fake-reviews-testimonials (accessed Nov 2026)

Frequently asked questions

Does Facebook read text in images on ads?

+

Meta has not said it runs OCR on ad images specifically. It has said, in an October 2026 engineering post, that its multimodal models read each ad's text and images and turn them into an embedding used to match ads to people. Text rendered on a static image is very likely part of what those models capture.

Is the 20% text rule still in effect on Facebook ads?

+

No. Meta removed the 20% text limit on ad images in September 2020. Image ads with more text are no longer capped or held back by that rule. Meta still suggests keeping on-image text concise, which is good practice anyway because most people see your ad on a small phone screen while scrolling.

Is a Meta Entity ID real?

+

Entity ID is a term used by agencies and tool vendors, not by Meta. No Meta page we found uses it. The closest confirmed idea is Meta's October 2026 statement that, in its ad tokenization system, similar items share tokens, and its guidance that too-similar creatives may be treated as the same and shown to the same audiences.

Does on-image text work like interest targeting?

+

Not directly. Meta's system predicts which ad each person is most likely to act on. Your on-image text is one input to how the ad is represented, which changes who the system predicts it is relevant for. It does not select an audience the way an interest or keyword setting does.

What does creative is the new targeting mean?

+

It is practitioner shorthand for Meta's own position. In April 2025 Meta called creative diversification the best lever to find the most relevant audiences. With broad targeting, the message, persona and visual in each ad do much of the audience segmentation that interest targeting used to do.

Keep reading

Or browse all articles.

Put this into practice today

4,000+ static ad templates modeled on proven winners, organized by niche and format. Open any of them in Canva and launch in minutes.