What is a multimodal model, in plain English?
The first generation of chat AI read and wrote text, and nothing else. A multimodal model adds other senses. Show it a picture and it can describe it; play it speech and it can follow it; give it both and a question and it answers from all three. The models behind ChatGPT, Gemini and Claude have taken images since 2023 and 2024, and voice since soon after.
For an owner, the practical meaning is this. Before, to get advice on your Google Ads you had to describe the account in words, and you described the symptoms. Now the AI can look at the account and see the cause. The description step, where most of the useful detail was lost, is gone.
What changed for an owner?
Two things follow. First, advice got specific. An AI that can see the campaign settings, the landing page and the search terms report gives you the answer for your account, not the answer for everyone. Second, the way you talk to it changed. Voice is a mode too, so the conversation can happen out loud while you work, with the screen as the shared thing you are both looking at.
Three limits are worth knowing. A multimodal model reads a screenshot, so it sees what is on the screen at that moment and nothing underneath it. It reads numbers off an image, so a nightly read of the account itself is more reliable than a glance at a dashboard, which is why CoraVision does both. And a model that can see is not a model that can act: looking at a Save button and clicking it are different abilities, and a product can have the first without the second.
What does it read off Ads Manager?
An owner opens Meta Ads Manager during a session. The screen shows an ad set with 50,000 impressions, a reach of 12,500 and a frequency of 4.
A text-only model, told "my Meta ad has 50,000 impressions and no sales", would ask questions it cannot answer. A multimodal model reads the three numbers from the image, works out that 12,500 people have each seen the ad four times, says so out loud, and asks whether the audience is new people or past visitors, because the answer decides whether 4 is a problem. The owner says new people. The advice is a second creative or a wider radius, pointed at on the screen.
How does CoraVision use one?
CoraVision runs on a multimodal model, which is how a spoken conversation and a look at your screen happen in the same breath. Cora sees your screen as still images during a session you start, and the images are not saved to your account. The conversation is spoken. Between sessions, CoraVision reads Google Ads, Meta and Google Analytics 4 directly, every night, which is the reliable read, and never changes anything in them.
Questions
Is a multimodal model the same as vision AI?
Vision AI is the older, narrower idea: software that understands images. A multimodal model handles images as one input among several and can talk about them. Most of what is called vision AI today runs on a multimodal model.
Does CoraVision use one?
Yes. It is how Cora can look at a still image of your screen and talk with you about it in the same breath.
Does the model see my screen all the time?
In CoraVision, no. It sees still images during a session you start, and they are not saved to your account.