Audiences can now talk back to the algorithm!
For most of the history of social media, algorithms have had to guess what we want.
You watched a video until the end.
You liked a post.
You clicked on another.
You followed an account.
You paused for slightly longer on a particular Reel.
From thousands of small behavioural signals like these, recommendation systems built an increasingly detailed picture of what we were interested in.
The relationship was fairly simple:
We behaved. The algorithm interpreted.
But that relationship is beginning to change quite dramatically.
Facebook is now rolling out a feature that lets users tell its recommendation algorithm, in their own words, what they want to see more or less of.
Not through a series of categories.
Not just by clicking “Interested” or “Not interested.”
You can actually explain it. Talk to it!
Meta says its large language models can interpret the interests, context and intent behind that language and adapt Feed and Reels recommendations in real time.
Audiences can now quite literally talk back to the algorithm.
And I think that is much more interesting than the feature itself.
Because for marketers, it hints at a shift from a world built almost entirely around inferred intent towards one where declared intent becomes part of the recommendation system too.
Algorithms have always watched what we do
Recommendation algorithms became extraordinarily powerful because behaviour reveals things we don't necessarily tell platforms explicitly.
Someone might never say:
“I'm currently interested in renovating a Victorian house.”
But if they suddenly start watching kitchen renovation videos, saving interior-design posts, following architects and clicking on furniture recommendations, the platform doesn't need them to say it.
Their behaviour says it for them.
This is what digital platforms have become incredibly good at.
They observe thousands of signals and try to predict:
What is this person likely to want next?
And from a marketing perspective, that changed everything.
Instead of primarily choosing audiences ourselves, increasingly we create content or advertising and let algorithms find people likely to respond to it.
The marketer provides the inputs.
The machine increasingly handles the matching.
But there has always been a limitation.
Behaviour is evidence of interest.
It isn't necessarily an explanation of interest.
Watching something doesn't always mean you want more of it
Anyone who has spent time on TikTok, Instagram, YouTube or Facebook has probably experienced this.
You watch one video about something completely random.
Maybe it's an obscure historical event.
Maybe it's a particular diet.
Maybe it's videos of people restoring rusty tools.
Suddenly your feed decides:
Excellent. This is who you are now.
Here are another 47.
From the algorithm's perspective, the logic isn't unreasonable.
Your behaviour indicated interest.
But human intent is more complicated.
Sometimes we watch something because we genuinely care about the subject.
Sometimes because it surprised us.
Sometimes because we're researching something for work.
Sometimes because someone sent it to us.
Sometimes because we're annoyed by it.
Sometimes because we're simply curious for thirty seconds.
The behavioural signal looks similar.
The intention behind it can be completely different.
That is where natural language becomes interesting.
Instead of forcing the system to reverse-engineer intention from behaviour alone, the user can provide context.
“I'm interested in AI, but I don't want AI-generated art.”
“Show me more running content, but less content about marathon training.”
“I'm travelling to Rome next week. Show me restaurants and architecture for the next few days.”
Those are much richer instructions than a like.
This isn't Facebook's first step in this direction
Meta has been gradually experimenting with this idea across its products.
Threads introduced Dear Algo earlier this year, allowing users to write things like “Dear Algo, show me more posts about podcasts”, temporarily adjusting what appears in their feed. It later expanded this with Your Algo, letting people privately specify topics they want more or less of and choose how long those preferences should apply.
Instagram has also introduced tools allowing people to customise the topics influencing recommendations.
Facebook's latest version pushes the concept further by letting people give more detailed, freeform feedback and having recommendations respond within the same session.
Taken individually, these can look like product updates.
Taken together, I think they point towards something bigger.
Recommendation is becoming conversational.
From inferred intent to declared intent
For marketers, I think there is a useful distinction here.
Platforms are increasingly able to understand three different kinds of signals.
Behavioural intent
What you do.
What you watch.
What you click.
What you buy.
What you ignore.
Historically, this has been the dominant signal behind personalised recommendations.
Declared intent
What you explicitly say you want.
“Show me more independent films.”
“Less football.”
“More technical AI content.”
This is different because there is much less interpretation required.
The user has actually told the system what they are interested in.
And for the first time the gap between Behaviour and the (albeit declared) Reason behind it seems to be closing.
Contextual intent
What you care about right now.
This could be temporary.
You might spend three weeks researching mortgages without wanting your online identity to permanently become “person who loves mortgage content”.
You may suddenly care about Formula 1 because you're attending a race next weekend.
You might research baby products because you're buying something for someone else.
Natural language gives recommendation systems a much better way of understanding that temporary context.
And that could make audiences considerably more dynamic.
But what we say and what we do aren't always the same
There is an important complication here.
People are not necessarily reliable narrators of their own preferences.
Ask someone what kind of content they want to consume and they may say:
More educational content.
More industry analysis.
More documentaries.
More thoughtful journalism.
Then look at their behaviour at 11:45pm and discover they've watched 26 videos of dogs falling off sofas.
There is a difference between stated preference and revealed preference.
Recommendation algorithms know this very well.
Which is why I don't think conversational algorithms will replace behavioural ones.
Meta isn't saying that either.
Its new system combines freeform user input with the existing information it has about what people watch, like, comment on and interact with.
So the future probably isn't:
“Tell the algorithm what you want.”
It is:
“Tell the algorithm what you want, while it continues watching what you actually respond to.”
That combination could be significantly more powerful.
And slightly uncomfortable, depending on how honest your late-night scrolling habits are.
So what changes for us marketers?
This is where I think things get genuinely interesting.
For years, we have spent enormous amounts of energy understanding which signals make algorithms distribute content.
Watch time.
Likes.
Shares.
Comments.
Posting frequency.
Retention.
Hooks.
Keywords.
Trending sounds.
All of those things still matter.
But if recommendation systems become better at understanding a user's semantic intent, another question becomes increasingly important:
Does the machine clearly understand what your content is actually about?
Imagine someone tells their feed:
“I want more practical advice for running a small marketing agency.”
The recommendation system now has an explicit request.
It needs to search through an enormous content universe and decide which posts genuinely satisfy it.
Suddenly, being clear about what your content is about matters enormously.
Not just to humans.
To the system matching humans with content.
Vague content may become harder to recommend
This creates an interesting problem for brands (or an annoying one depending your view on things).
A lot of corporate content is remarkably vague.
“Five ways to unlock growth.”
“The future of customer experience.”
“How businesses can succeed in the age of AI.”
They sound vaguely relevant to almost everyone.
Which means they are specifically relevant to almost nobody.
In a world where algorithms mainly optimise around engagement, broad content can sometimes work because strong hooks create enough behavioural signals.
But conversational recommendation creates a different challenge.
If a person tells an algorithm:
“Show me practical advice about improving profitability at a 30-person marketing agency.”
What exactly is your generic “Unlock Your Growth Potential” post giving the system to work with?
The clearer the audience's intent becomes, the clearer content may need to become in response.
Topic ownership could become more valuable
This also makes me think differently about content consistency altogether.
Brands have historically been encouraged to build authority around a few topics because it helps audiences understand what they stand for.
But it may increasingly help recommendation systems understand them too.
Imagine a company publishes consistently around:
marketing measurement;
agency profitability;
paid media;
and reporting automation.
Over time, there is a very strong semantic picture of where that brand belongs.
Now imagine another company publishes:
AI one day;
leadership the next;
a motivational quote on Wednesday;
something about sustainability on Thursday;
and a meme on Friday.
One may generate plenty of impressions.
The other is much easier to classify.
If users increasingly tell algorithms the subjects they actively want, being strongly associated with a subject could become more useful than being vaguely relevant to many of them.
Engagement bait becomes a weaker long-term strategy
There is another implication.
A lot of social content is optimised to trigger behaviour rather than satisfy intent.
The exaggerated hook.
The controversial statement.
The deliberately incomplete explanation.
The carousel that makes you swipe twelve times to discover something that could have been written in two sentences.
These techniques can create strong behavioural signals.
But if recommendation systems increasingly understand what users explicitly say they want, there may be a growing distinction between:
content that earns engagement
and
content that satisfies intent.
The two obviously overlap.
Good content can do both.
But they aren't identical.
A user may click a sensational post and still tell the algorithm:
“Please show me less content like this.”
That's a fascinating signal.
The click says yes.
The person says no.
Which one should the algorithm believe?
Increasingly, it can listen to both.
Audience targeting may become more fluid
There is also a broader implication for how marketers think about audiences.
We've traditionally loved putting people into relatively stable boxes.
Marketing director.
Small-business owner.
Running enthusiast.
Luxury traveller.
Tech buyer.
But human interest isn't static.
At 9am, someone might be a CFO researching accounting software.
At 6pm, they're a parent looking for somewhere to take their children.
On Friday, they're planning a weekend away.
Next month, they're buying a house.
The same person temporarily becomes a completely different audience depending on what they are trying to accomplish.
Conversational recommendation systems make it easier for users to express those temporary states directly.
That could move personalisation further away from:
Who is this person?
and towards:
What does this person want right now?
For marketers, that's a subtle but important difference.
This isn't really about Facebook
Facebook's new feature may work brilliantly.
Or users may barely touch it.
That's almost beside the point.
What interests me is the interface.
For decades, our relationship with algorithms has mostly been indirect.
We performed actions.
They watched.
They guessed.
And results where coming out of the pipeline.
However, now Large language models introduce another possibility.
We can explain ourselves.
And once machines can understand natural-language intent well enough, it seems unlikely that this idea stops at social feeds.
Search.
Shopping.
Streaming.
Travel.
News.
Marketplaces.
Advertising.
Any system trying to decide what should this person see next? can potentially benefit from simply asking the person.
Of course, what we say will never completely replace what we do.
Sometimes our behaviour is more honest.
But the combination of the two changes the relationship.
The algorithm isn't just watching anymore
Marketing has spent years trying to understand algorithms.
How do we rank?
How do we get recommended?
What makes the platform distribute this?
Which signals matter?
Perhaps increasingly, the audience gets to participate in that conversation too.
The recommendation engine still watches what people do.
But now it can also ask what they mean.
And people can answer.
For marketers, that creates a fairly simple challenge:
If someone could tell an algorithm exactly what they wanted to see…
would your content be an obvious answer?
This new feature has the potential to disrupt heavily the way we approach our marketing campaign strategy philosophy currently.
Things seem to be getting more fluid by the month and with LLM's into the mix things are going to be changing every so often.
Keeping up with the changes, staying true to one or two topics that your business is all about and having a clear positioning and messaging that you signal out in the web in a similar (if not the same) way across different channel seem to be the only available tools marketers are going to have to rely on for the time being.
And with Facebook making the first step towards that direction who do you think is going to follow?