Put the evidence next to the decision

For the software comparison, a link to the pricing page is a start. A better result puts the regular price next to the introductory rate and says when the price changes. If there’s a minimum seat count or an annual commitment, include it next to the figure, too. If a vendor doesn’t publish a price, the table should say so. A blank cell with a reason is more useful than a confident number somebody has to unwind later.

Try three comparisons before rolling this into a workflow. Give one pair of vendors straightforward public pricing. Include an introductory offer with a price increase in the second. For the third, pick a vendor that makes you request a quote. Check whether the assistant distinguishes those cases and whether a coworker can verify the deciding detail without rerunning the search or reading the whole page.

That last part matters if the comparison is going into an email or spreadsheet. A citation that only exists inside the chat is easy to lose when someone copies the recommendation. Keep the source and the condition attached to the figure wherever the answer ends up.

Keep an eye on how much work the search does

DeepSeek bills successful searches and page reads separately from model tokens. It also offers standalone search and fetch endpoints if your application needs to choose the queries and pages itself. One approach gives the model room to follow a lead; the other gives your app more direct control over what gets looked up.

Compare the cost and time of getting an answer someone can actually use, not just the cost of producing a table. If two supplied pricing links tell you everything, a broad search may add little. If the vendor has buried a crucial limit in separate terms, another page read may be the most valuable part of the job.

Let the assistant do the digging. Make the detail that decides the answer easy to check.