How to Optimise Content for Social Search on Meta.

To optimise content for social search on Meta, write captions that state plainly what the post is about, in the words a real person would use to ask for it.

That is no longer a guess about how ranking works.

On Meta’s second quarter 2026 earnings call, CFO Susan Li said every public Reels and Feed post on Instagram is now automatically processed through a large language model and analysed “across dimensions from topics to tone”, with those signals passed downstream into ranking, recommendations and content policy enforcement.

A system that reads what your post is about rewards clarity and gets almost nothing from a block of hashtags. That changes what a good caption looks like, and it changes who in your team should be writing one.

What Meta actually said about reading every post

The line came in the prepared remarks of the call on 29 July 2026, in the section on content recommendations rather than the section on advertising.

Li said Meta had reached a milestone earlier in the year of every public Reels and Feed post on Instagram being automatically processed through a large language model, analysed across dimensions from topics to tone, and that the company is “working towards including more surfaces on Facebook as well”. Those signals, she said, are then passed to downstream applications across ranking, recommendations and content policy enforcement.

Read the scope carefully. Instagram. Public posts. Reels and Feed.

Facebook is described as work in progress, not as done.

What Meta did not say

Meta did not say it ranks posts by keyword. It said a model reads each post and produces signals that other systems then use.

That distinction leads to two different captions. Keyword ranking rewards getting the term in. Content understanding rewards being unmistakably about one thing.

Meta also did not say the model reads only the caption, did not say a well classified post gets more reach, and put no number on any of it. Ranking is one of three named uses, sitting alongside recommendations and policy enforcement.

Anyone who tells you they know the weighting is guessing.

Search matches words, recommendations read meaning

There are two systems here and brands keep collapsing them into one.

Instagram Search is the older of the two and it works on text. Instagram’s own explanation of it, from Adam Mosseri in 2021, says the platform tries to match what you type with relevant usernames, bios, captions, hashtags and places, and his advice to creators was to put keywords and hashtags in the caption rather than the comments. That still holds.

The recommendation layer is the new one and it works on meaning. Search serves people who already went looking. Recommendations decide whether anyone who was not looking ever sees you, and on Meta that is most of your audience.

The caption feeds both, which is why the shift to keyword-driven social strategies only ever described half the job.

Stop stuffing keywords, start stating the subject

A stuffed caption and a clear caption often contain the same words. Only one of them reads as a sentence.

Stuffed: “marathon running shoes, best running shoes, marathon training, running tips, shoe review, runners of Instagram”.

Clear: “We ran 40km in the new carbon plate marathon shoe to work out whether it is worth the price.”

The second names the subject, the format, the question and the verdict on offer. A model asked what this post is about has exactly one answer. A person scrolling has one reason to stop.

The rule we use is simple. Write the sentence you would say out loud if a friend asked what the video was. That sentence is usually the caption, and it usually contains the search terms anyway, without anyone having to force them in.

The system reads the whole post, not just the caption

Meta’s statement is about the post, not the caption. Li also said Meta had begun using its Muse family of models for content understanding across signals like video topic classification and summarisation.

So the caption is one input among several. On-screen text, spoken words, the opening frames and the cover image are all carrying the same job.

The test is easy to run. Mute the video, hide the caption, watch three seconds. Could you say what it is? If not, the file itself is ambiguous, and no caption rescues an ambiguous file.

This is ordinary platform optimisation work, and for brands already posting consistently it is usually where the gap sits.

Tone is a measured dimension now

Topics to tone. Tone is the half that gets skipped, and it is the half with consequences, because the same signals feed content policy enforcement.

So the register of your writing is no longer only a brand-voice decision. It is an input to the system that decides how far a post travels and whether it is restricted.

We are not claiming a sharp caption gets suppressed, and Meta did not say that either. The claim is narrower and more useful: your tone is now legible to the machine in a way it was not when the system was mostly reading hashtags.

If reach currently depends on bait, that is a risk you cannot see in your analytics.

The read happens at publish, so the caption has to be right first time

On the same call, Li said Meta’s largest ranking models can now identify high-quality new Reels at creation, and that over half of all recommended content on Instagram Feed is less than one day old, more than double a year ago.

That is an operations point, not a copywriting one.

The caption you publish is the caption the system reads while the post still has a chance. Editing it two days later, once someone notices it underperformed, is not a fix. The distribution decision has already happened.

Which means caption writing belongs in the brief and in the review, not in the ninety seconds before someone hits publish.

Your audience is describing what it wants in plain language too

The demand side moved in the same quarter. Instagram users can open a “Your Algo” page and write natural-language prompts to tune their own recommendations. Facebook launched “Shape Your Feed”, which Li said is showing over 80% retention among the people who engage with it.

Both sides of the match are sentences now. Someone asks for less of one thing and more of another, and a model that has read your post decides whether you are the more.

You cannot write for that with a hashtag block. You write for it by being obviously and repeatedly about a specific thing.

What is supported by Meta, and what is our judgement

Two lists, because a lot of advice on this subject will blur them.

From Meta’s own statement:

• Every public Reels and Feed post on Instagram is processed through a large language model.
• Posts are analysed across dimensions from topics to tone.
• Those signals feed ranking, recommendations and content policy enforcement.
• More Facebook surfaces are planned, not shipped.

Our judgement, not Meta’s:

• Plain, specific captions classify more cleanly than stuffed ones.
• An ambiguous video needs an unambiguous caption more than a clear video does.
• Topic consistency compounds, because the same reading happens on every post you publish.
• Hashtags are now a small part of a caption’s job rather than most of it.

The second list is what we would change on Monday. It is reasoning from how classification works, and we would rather label it as reasoning than dress it up as a rule Meta published.

What good looks like

Imagine a running brand posting four times a week. Every video opens on the thing itself. Every caption names the shoe, the distance and the question being answered. Nothing in the file depends on a hashtag to explain it.

Across a month, every post says the same thing about what the account is for. A model classifying topic reaches the same answer sixteen times in a row.

Now imagine the same brand with clever captions, in-joke openers and a hashtag block underneath. The videos might be better made. The account is harder to place, and this is a platform that places accounts constantly.

Good, here, looks slightly boring on the page and very precise underneath it.

How NBK thinks about social search on Meta

We treat this as an operations problem, because that is where it actually breaks. Most brands are perfectly capable of writing a clear caption. What they lack is a step in the process where someone owns it.

So the caption gets written last, by whoever happens to be publishing, with no view of the other posts going out that week. Multiply that by a year and the account has no legible subject.

Fix the step and the rest follows. The subject gets named in the brief, the caption gets reviewed like the edit gets reviewed, and the account holds a line. It is also why topical consistency beats posting volume on a platform that now reads every post you publish.

Next step

If your social output feels busy but not effective, an audit is the quickest way to find out whether the problem is the content or the system producing it.

On Meta specifically, there is a version you can run yourself this afternoon. Take your last twenty posts, read only the captions, and write down what each one is about. If you cannot, neither can the model.