TL;dr: Fill with Gemini in Google Sheets now puts a source citation on every AI-filled cell, up to 3,000 cells per request, and Google says it is up to 55 times faster than typing the same data by hand. That citation is the audit trail AI-filled models were missing, because until now an AI-filled cell was just a number with no provenance, and a number with no provenance does not survive a close review. The catch is that a cited source is not a verified number, so the checking ritual stays, and this post walks through it with a vendor price list you can copy.

What changed is small on the surface and large in practice. Google started rolling this out on October 20, and announced it on October 28, as part of a wider speed and consistency update to Fill with Gemini. When you ask Gemini to fill a range, each filled cell now carries the web source it came from, so you can click a cell and see where its value originated instead of taking the model's word for it. For finance work, provenance is the whole game, because a model that fills a price list is doing the same job as a junior analyst, and you would never accept a junior analyst's numbers without asking where they came from.

It helps to be precise about what the citation actually shows, because that tells you what it does not show. The citation answers one question, which is where the value came from. It does not tell you when the source page was last updated, whether the number on that page is still current, or whether the model read the right row of the right table. A cited price from a page last updated in March is still a March price, and a cited price from the wrong variant of a product is still the wrong price, just with a footnote attached. So the citation turns an unverifiable number into a checkable one, and the checking is still your job.

Setting up your verification ritual: first, make a copy of the template below, because the template itself is read only and your copy is where the work lives. Second, go to the Source-audit checklist tab and work through it in order, because the order matters. Start with a sample, not all 3,000 cells, and pick the sample the way an auditor would, meaning the largest values, the ones that moved most against the prior period, and a random handful on top. Third, for each sampled cell, open the source and find the number on the page, then confirm four things before you mark the row verified: the number matches, the date on the page is current, the units match your column, and the context is right, meaning the price is for the item and variant you actually buy. Fourth, tie the totals back, because a total that moves more than you can explain is a flag to dig into, not a number to accept. Fifth, mark each row in the checklist, and treat a row as verified only when all four checks pass, because a half checked row has a way of becoming a trusted row by the time the model is presented.

Worked example, using the numbers already in the template: a 12 line vendor price list for office supplies, with a prior quarter total of 5,206 dollars. The AI-filled total came in at 6,754 dollars, up about 30 percent, which is the kind of jump that should stop you, because office supplies do not rise 30 percent in a quarter. The sample found three problems. Vendor F, copy paper, had a unit error: the model filled a case price of 180 dollars into a unit price column, and on a quantity of 12 that overstated the line by 1,440 dollars. Vendor C, ink cartridges, had a stale source: the cited page was last updated in March, the filled price was 85 dollars, and the current price on the vendor's site is 105 dollars, so that line was understated by 200 dollars on a quantity of 10. Vendor K, specialty labels, had a paywalled source, so the cited page could not be opened at all, which means that 310 dollar line stays marked unverified no matter how clean it looks. After corrections the verified total is 5,514 dollars, up about 6 percent on the prior quarter, which is a believable number, with exactly one row still flagged. That is the ritual working as intended: a 30 percent jump that meant nothing became a 6 percent increase you can defend, plus one honest gap.

Where it still breaks, beyond the example, comes down to three failure modes worth naming so you recognise them when they show up. Stale sources are the first, because a citation proves where a number came from, not that it is current, and prices, rates, and terms all drift. Wrong context is the second, because the model can pull the right number from the wrong table, like a list price instead of a contract price, a US price instead of a Canadian one, or a price for a different size or variant, and the citation will look perfectly respectable while the number is wrong. Paywalled data is the third, because a citation pointing at a page behind a login is a citation you cannot check, and an uncheckable cell is an unverified cell, full stop.

When to trust it versus when to fill by hand follows from that. Trust it with sampling for large, stable lists where the sources are public and current, like published list prices, standard catalog items, or reference data that changes slowly. Verify every cell, not a sample, when the number moves money directly, like a payment run or an accrual, and when the source sits behind a login. And fill by hand when the data is contractual or negotiated, because a contract price does not live on a public web page in the first place, so there is nothing for the model to cite. The rough rule is simple: the more a number matters and the less public its source, the more of the ritual you do.

The citation turns AI fill from a black box into a worksheet you can audit, which is exactly what finance work needed from this feature. Run the ritual, keep the checklist, and the speed claim starts to be real instead of just fast, because a filled range you can defend is worth far more than a filled range you merely received.

Template: template: AI-fill source-audit checklist (Google Sheets, make a copy to use it)

- Francois