Mozilla has fixed one of the stranger problems with Firefox’s AI-powered Shake to Summarize feature: it was summarizing recipes too much.

That sounds almost like a joke, but Mozilla’s testing found that its general-purpose AI prompt could turn a perfectly good recipe into a vague description of the dish. A lentil soup recipe, for example, came back as a description of the ingredients and cooking approach rather than something a person could actually cook.

Mozilla’s solution was to create a dedicated recipe prompt that preserves the information that matters: servings, preparation and cooking times, ingredients, instructions, tips and nutrition information when available. Mozilla says the revised system was more than twice as likely to produce a complete and accurate recipe summary in its test set.

The fix works. But there’s a more interesting question hiding underneath Mozilla’s announcement: when does a recipe actually need an AI summary?

Because the answer isn’t simply “never.”

In fact, Firefox’s optional, user-triggered approach makes the feature considerably more useful than it might sound at first.

Recipe websites already solved part of this problem

Visit a major recipe website today and there’s a good chance you’ll find a dedicated recipe card.

You know the thing: servings, prep time, cooking time, ingredients, instructions and sometimes nutrition information, all neatly separated from the story surrounding the recipe.

There’s a reason publishers do this. Recipe websites have spent years making their content easier for both humans and machines to understand. Structured data such as Google’s Recipe markup can describe ingredients, instructions, preparation and cooking times, servings, nutrition and other recipe information.

Mozilla is now using that kind of existing webpage structure inside Firefox.

Its recipe summarization system checks structured data embedded in a webpage to help determine what type of page it is. If Firefox identifies the page as a recipe, it can route the content to a specialized recipe prompt rather than treating it like a generic article. Mozilla says this approach avoids using another AI pass just to classify the page.

And that creates a funny little bit of technological circularity.

The website has already structured the recipe.

Firefox recognizes that structure.

Firefox then asks an AI model to produce a mobile-friendly version of information the website may already have neatly presented.

That isn’t necessarily bad. But it isn’t exactly revolutionary either.

If I’m already looking at a recipe card with the ingredients and instructions right in front of me, I probably don’t need to shake my phone and ask Firefox to tell me what I just read.

The important thing is that I have already visited the website.

But what if the recipe website is a mess?

Now flip the situation around.

You search for a recipe and land on a small website you’ve never seen before.

There isn’t a polished recipe card.

There may not be useful structured data.

The recipe might simply be written as a long article. You scroll. And scroll. And scroll some more.

There’s a story about how the author’s grandmother used to make the dish. Then an anecdote about a family vacation. Then a photograph. Then another paragraph.

And somewhere in all of that is the actual recipe.

Eventually, you’re thinking:

“I didn’t find anything useful here. Let’s shake it.”

That’s where Firefox’s feature starts making a lot more sense.

Mozilla specifically designed the recipe prompt to preserve the details needed to use the recipe, including ingredients, timings and key instructions. Its revised output can turn a messy page into something much closer to a compact recipe card.

For a website that already has a beautifully designed recipe card, that’s arguably redundant.

For a website that doesn’t, Firefox can potentially rescue the experience.

And that is a much stronger argument for the feature than simply saying “AI can summarize recipes.”

The difference is the website, not the AI

This is probably the most interesting takeaway from Mozilla’s experiment.

The usefulness of Shake to Summarize isn’t really binary. It depends on the page you’re looking at.

On a well-structured recipe website, the feature may offer little that the publisher hasn’t already provided.

On an unstructured recipe article, it can turn a frustrating page into something much easier to use.

That’s especially relevant because Mozilla isn’t requiring everyone to use it. Shake to Summarize is an optional tool that users can choose to activate after opening a webpage.

So the workflow becomes surprisingly reasonable:

“I can see the recipe card. I’m good.”

No AI required.

Or:

“Where on Earth is the recipe on this page?”

Shake.

That’s a useful distinction.

There’s still a publisher question

There is a legitimate publisher question here, but it is more nuanced than saying Firefox is automatically taking traffic away from recipe websites.

The visitor has already arrived.

That’s important.

The potential issue is what happens after that visit.

If Firefox gives someone exactly what they need from a poorly structured recipe page, the user may stop reading sooner. They may not browse the rest of the article, look at another recipe or interact with the surrounding content.

For a publisher, that could matter.

Shake to Summarize operates at the page level. The user is already there, and the user chooses whether to activate it.

That makes the feature fundamentally different from a system that gives users an answer before they ever visit the publisher’s website.

Mozilla’s recipe fix is also a lesson in good AI design

There’s another part of Mozilla’s announcement worth giving credit for.

The company didn’t simply throw a larger model at the problem.

It discovered that the definition of a “good summary” changes depending on the type of content. A news article needs its important facts. A review needs its key findings. A recipe is different.

For a recipe, the “summary” may need to preserve much of the original recipe itself.

Mozilla responded by separating its prompts by content type and routing recipe pages to a dedicated recipe prompt. It says the experience reinforced the idea that reducing ambiguity and narrowing the task can improve results without necessarily requiring a larger model.

That’s a useful lesson beyond Firefox.

Generative AI doesn’t always need to replace an existing system. Sometimes it works best as a layer that fills in the gaps where existing systems aren’t enough.

Recipes might be a perfect example.

So, should Firefox keep Shake to Summarize?

For now, I wouldn’t frame the feature as something Firefox shouldn’t have built.

It’s optional, it’s activated after the user reaches a webpage, and there are plenty of pages where it could be genuinely useful.

But Mozilla’s own experiment also highlights why context matters.

If a publisher has already built a good recipe card and supplied structured data, Firefox is arguably adding another layer to a problem the web has already solved.

If the publisher hasn’t done that work, Firefox can give the visitor a much easier way to extract useful information from the page.

That’s what makes Shake to Summarize interesting. It’s not equally useful everywhere. On some recipe sites, it may feel redundant. On others, it could be exactly the feature a frustrated visitor needs.

So maybe Firefox doesn’t need to shake it off after all.

Maybe it just needs to know when to shake it up.