A client I build sites for sent over their meeting materials as a single PDF late one Friday. Scanned paper and screen captures, bound together into something with a lot of pages and not much structure.
I dropped it straight into the Gemini app, hoping to pull the key points out before the weekend. What came back was not the summary I had in mind.
So I started splitting. Half, then half again, and at roughly a quarter of the original the answers finally made sense. That detour cost me about half an hour — and when I went back to the help pages afterwards, I found that splitting had barely been the point at all.
The same rejection can come from three different walls
The Gemini app has three separate messages for turning a file away, and they mean genuinely different things.
| Message | What you actually hit | Does splitting help? |
|---|---|---|
| Delete data to upload file | Storage inside Gemini Apps — your past chats and uploaded files. It is counted separately from Google Drive | No |
| You've reached your limit for chats with files | A rolling rate limit on how often you can upload and analyse. It resets on its own | No, and each retry spends more of it |
| Your uploads may be too large for the best results | Reading capacity — the context window. The file was accepted; the model simply cannot hold all of it at once | Yes, but something else helps more |
Mine was the third one. The file had gone through perfectly well. From where I was sitting, though, all three look identical: you get a reply that misses things, and you assume the file was too big.
One rejection, three repairs. Reading the message to the end is the cheapest step in this whole process.
Fitting in is not the same as being read
The entry limits are more generous than most people expect. Up to 10 files in one prompt, up to 2 GB for a video, and up to 100 MB for everything else. A 100 MB PDF is not going to be turned away at the door.
Reading capacity is where the plans diverge sharply.
| Plan | Context window | Google's own comparison |
|---|---|---|
| No Google AI plan | 32,000 tokens | Short documents only |
| Google AI Plus | 128,000 tokens | Four times that |
| Google AI Pro / Ultra | 1,000,000 tokens | Around 1,500 pages of text, or 30,000 lines of code |
Google only publishes the page comparison for the one-million figure, so scaling it down puts a 32,000-token window at a few dozen pages of text. A door that admits 100 MB, and a window that reads a few dozen pages, sitting inside the same app.
What fits through the door and what fits through the window are two different limits. The reason my repeated halving eventually worked was accidental: I had cut the document down to something the window happened to hold.
There is one more ordering that surprises people. Deep Think, the Ultra-only mode that reasons the hardest, runs with a 192,000-token window. Switch up to the more capable feature and your reading capacity narrows. Hand it a thick document and you will see more gaps, not fewer.
If you are still deciding which plan to pay for, Gemini Free vs Pro vs Ultra - Choosing the Right Plan for You (2026) covers that side.
Five things I trim before I split anything
If the document will not fit the window, the honest move is to make it smaller. This is the order I work through on client material, most effective first.
- Drop the cover, table of contents, revision history, blank pages, and colophon. Pages that exist for presentation carry no answers. On that client PDF, this alone removed about a fifth of it.
- Remove chapters that have nothing to do with today's question. Appendices and reference material go into a separate file rather than the bin, so I can hand them over later if they turn out to matter.
- Thin out pages that repeat the same high-resolution figure. Scanned pages consume far more capacity than pages of plain text, which is why an image-heavy document fills the window sooner than its page count suggests.
- Send text as text wherever you can. Specs and meeting notes I already hold in plain form go in directly, not flattened into a PDF first. They travel much lighter.
- Stop asking it to read everything, and fix the question first. Reducing the number of things you want answered stabilises the result more reliably than reducing pages does.
In my case, keeping only the meeting notes and the specification chapters was enough to get through without splitting at all. What I had been doing that first evening was dividing the weight evenly without ever removing any of it.
When you do split, decide how first
Some documents really are too large, and then splitting is unavoidable. The one rule worth settling in advance is not to divide by page count.
Cut on chapter boundaries and keep one subject per file. Even division tends to strand a single topic across two files, so neither half answers the question on its own.
More often I hand over only the table of contents first and let it point me at the right section.
Looking only at this table of contents, name up to three chapters
that could answer the question below. If you do not think any
chapter can answer it, say so in one line and explain why.
Question: where is the line between features shipping in the first
release and features deferred to the next one?Then I send just those chapters, and the window stops being a concern. As a side effect, I get oriented in a document I did not write — which, as an indie developer reading other people's specifications most weeks, I have come to value more than the token savings.
If you handle 200-page documents regularly, Reading a 200-Page Contract with Gemini 2.5 Pro — Five Techniques That Move Long-Context Analysis to Production Quality moves this toward a chapter-by-chapter design with cross-checking.
The order I check things in when I hit a limit
If it was storage or the rate limit, trimming changes nothing. Having an order saves a round trip.
- Open your usage limits in settings. The Gemini app shows them under Settings. Limits refresh every five hours, with a weekly ceiling above that, so this screen tells you immediately whether waiting will fix it.
- If it was the storage message, clear old chats from Activity. Delete past chats that carried files. It takes several minutes to register, so if you retry straight away and see the same message, wait a little and try once more.
- Work and school accounts cannot clear this themselves. Only a Workspace administrator can delete Gemini Apps activity for those accounts. Following the personal-account steps will not get you anywhere, so ask your administrator early.
- If it is none of those, go back to trimming. Google's help does suggest re-uploading when you get an analysis error, so try that once; if nothing changes, shrink the document instead.
The same triage shape works elsewhere in Gemini. For a case where a limit turned out to be a file-location problem, see When the Sheets AI Function Won't Generate, Suspect the File's Location Before the Limit. If you are feeding documents from Drive and getting stale text back, I edited the Drive doc, asked Gemini again, and it quoted the old sentence back covers that one.
One thing to try on your thickest file today
Open the PDF with the most pages on your machine right now, and do only step one from the list above — count the cover, contents, revision history, and blank pages.
When I counted, close to a fifth of the document contributed nothing to any answer. Deciding whether to split can wait until after those pages are gone.
Both tables here come from Google's own help pages: Upload & analyze files in Gemini Apps and Gemini Apps limits & upgrades for Google AI subscribers. Google states plainly that these limits may change without notice, so those two pages are where I go back when something stops matching what I wrote here.
If this ordering saves you one round of pointless splitting on some Friday evening, it was worth writing down.