Skip to content

AI Became a Search Box, and Image Work Is Still a Spike

Two years of country-level ChatGPT usage data show writing falling from 35% of messages to 18%, search passing a quarter, and multimedia as the fastest grower at 7.8%. The shape matters more than the ranking: image work arrives in trend spikes, not habits.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
A search field and a blank sheet of paper side by side on a wooden desk, with a strip of small printed photographs fanned out beside them.
People ask, then they make something. The order is not the one most product decks assume.

Xataka published a piece this week that I keep reopening, not because the headline is surprising but because the numbers underneath it are. Spain's ChatGPT usage, tracked by OpenAI across two years, has quietly rearranged itself. In July 2024, writing was the dominant activity at roughly 35% of messages. Practical guidance sat at 24%. Information search was 19%.

Today writing is around 18%. It hit a low of 15% in August. Practical guidance is at 30%. Information search passed a quarter of all messages.

The thesis everyone built products on was that AI would write for us. In the country with the data, people mostly use it to look things up and to be walked through a task. Writing became the thing it does third.

The category that shrank is the one most products promise

If your landing page says the tool writes your copy, your captions, your emails, your scripts, you are selling into the shrinking bucket. That is not a claim that writing tools are dead. The absolute volume of messages has grown enormously while the share fell, so 18% of a much bigger pie can still be a bigger number than 35% of the old one. Shares move faster than totals, and reporters (me included) reach for shares because they are legible.

But the direction is still worth taking seriously, with one limit worth naming first: these are shares of messages, not a record of what any individual person did. Writing's share can fall while the number of people writing holds steady or even grows. The plausible reading is that a general chat model now competes with what people already had for writing tasks, a blank page and a decent search engine, and that some of those requests moved to the search side. The data cannot confirm the move. It can only make it the hypothesis worth testing.

There is a second read, and I think it is the more honest one, though the classifier does not support it as cleanly as I first assumed. Writing's own label set already covers the adjacent cases: proofreading and critiquing supplied text sits inside Writing, next to translation, summaries, personal messages and fiction. So the decline is not proofreading being filed somewhere else. What the taxonomy cannot show is intent. Some of the 18% may be people who used to ask for a document and now ask for a paragraph they will finish themselves, and the buckets cannot separate those two requests. The interaction may have shifted from "produce my document" to "answer me, then I'll write," but the data records the ask, not the reason behind it.

Multimedia is the fastest grower and the smallest category at once

OpenAI's own global report, published in August, puts multimedia at 7.8% of messages worldwide, growing faster than anything else since ChatGPT Images 2.0 shipped in April. In Brazil and Colombia, more than one message in ten is multimedia.

Both statements are true and they pull in opposite directions. Fastest growing means new demand. 7.8% means it is still, globally, a side activity. That combination has a specific name in product work: a spike category.

Xataka's timeline shows the spikes directly. Multimedia usage in Spain stays low, then jumps around March 2025, which the piece attributes to the Ghibli-style image craze that pushed ChatGPT image generation to absurd volumes, then settles, then lifts again after the Images 2.0 release. Trend event, surge, decay, next trend event.

This is the part I would put in front of anyone building a creative tool. Your acquisition comes from events. Your retention comes from whether the people who arrived during an event found a reason to come back when the trend was over. A spike is not a market. It is a set of people curious about a specific picture, and the share data says nothing about how many of them returned. Retention is the number to watch here, and this dataset cannot give it to you, because it counts messages rather than people.

Practical guidance at 30% is the quiet signal

The category that grew into the top spot is the boring one: practical guidance, which the report describes as personalized tutorials, advice on how to do a task, and requests for creative ideas. Thirty percent, in Spain, up from 24%.

Read that against what a template is. A template is not a prompt. It is a sequence: choose a subject, choose a look, keep the framing, change one thing. That is practical guidance with the model calls already wired in. Users are asking for exactly that shape in chat, in the least efficient possible way, one paragraph at a time.

So the honest framing for anything image-related is not "generate an image." It is "walk me from my photo to a finished asset, and tell me what to do at each step." That is also the shape that survives the trend cycle, because the guidance loop is reusable when the current trend dies. If you want the deeper mechanic of why step-shaped prompts beat one-shot prompts, the prompt decomposition breakdown holds up better now than it did when the usage data was just a hunch.

Work is "doing," home is "asking"

The global cut is cleaner than the country cut. At work, people are more than twice as likely to use ChatGPT to complete a task or produce an output. Outside work, asking dominates.

For anyone selling image work, that is a shortlist of where money actually moves. Product photos. Ad variants. Thumbnails. Deck visuals. Assets that someone needs by Thursday and would rather not brief a photographer for. None of those are curiosity-driven. They are task-shaped, deadline-shaped, and repeat weekly, which is the opposite of a trend spike and the reason the paid use of image tools does not look like the viral use of them.

The same report notes adoption spreading fast among people over 35, with the share of their messages up about 5 points year over year and more than 10 points in France and Czechia. That cohort is not chasing Ghibli portraits. They are making the thing their job needs, which is also why the productivity numbers and the booked-profit numbers keep pointing in different directions.

Where I would not trust the numbers

Two caveats worth stating, because this data gets quoted as if it were a census.

The classification is automated. An anonymized message gets read by software and dropped into one of seven buckets. The multimedia bucket is wider than its name suggests: it holds requests to create an image, requests to analyze an image, and requests for other media such as audio or video. So 7.8% is not the share of image generation. It includes people asking what is in a photo and people asking for a voiceover. The real image-generation number is smaller, by an unknown amount, and nobody outside OpenAI knows it.

The source is also the company being measured. Xataka makes the fair point that this is still one of the only windows into real behaviour rather than survey claims, which is worth something, but it is a window with a frame. And all of it describes individual use. The sample covers ChatGPT Free, Go, Plus and Pro, while enterprise traffic and Codex are left out, so OpenAI itself says the numbers likely understate business and technical work. That caveat lands directly on the previous section: the paid, task-shaped image work I just called the durable half is exactly the part this dataset is least able to see. A dedicated image tool also has a different usage mix, because nobody opens one to ask a question.

What to do with the shape

Three things I would take from this, in order of how much they change decisions.

Build the front door around events, and the second session around recipes. The trend brings a wave you did not earn. The recipe is the only thing that converts a wave into a user. If the only path from arrival to output requires reinventing the prompt, the wave leaves.

Treat guidance as the product, not the tutorial. The 30% category is people asking to be shown the steps. Most image tools hide the steps and expose a text box, which is the exact inverse of what the data says people want.

Do not read a shrinking category as a shrinking need. Writing's share fell while asking grew; the most plausible reading is that writing stopped being a destination and became part of asking — a hypothesis, since message shares cannot show what any individual switched to. The same thing is going to happen to image generation, probably soon: not a feature you go to, but a thing that happens inside the task you are already doing. Tools that insist on being a destination will keep renting traffic from trends. Tools that disappear into the workflow get the Wednesday product photo.

The trend is the loudest part of this data and the least durable. I would plan for the Wednesday.

Sources

Share this guide
X LinkedIn