A product spec spreadsheet becomes good chatbot knowledge when every column has a plain header with its unit, each row describes one product, and the file is saved as a comma-separated CSV with no line breaks inside cells. Upload it and each row turns into a line of text such as "Model: Trail 40 | Weight (kg): 1.2 | Waterproof: Yes", which the bot can find and quote. The layout decisions you make before export decide how well that works.
What happens to a CSV when you upload it
It is worth knowing the mechanics, because every layout tip below follows from them.
SpideyChat reads an uploaded CSV in one of three ways, based on its first row:
- Q&A mode. If the header row has a column called question (or q, prompt or faq) and one called answer (or a, response or reply), each row becomes a question and answer pair.
- Two-column mode. If the file has exactly two columns and the headers are not recognised, every row, including the first, is also read as a question and its answer.
- Table mode. Anything else is treated as a table. Each row becomes one line where every value is labelled with its column header, separated by vertical bars.
That text is then split into passages of about a thousand characters and stored. When a shopper asks something, the bot pulls the five passages closest in meaning to the question and answers only from them.
So a spec sheet needs to land in table mode, with headers that make each row understandable on its own.
Laying out the spreadsheet
One product per row. Each row should be a complete description of one item or one variant. If colours or sizes differ in specs, give each variant its own row.
Put the product name in the first column. Every row line then starts with "Product: ..." or "Model: ...", so the passage the bot finds always says which product it is about.
Write headers a customer would understand. The header is repeated in front of every value, so "Wt" or "SKU_DIM_H" turns into noise. "Weight (kg)", "Height (cm)" and "Battery life (hours)" read as plain facts. Shoppers ask "how heavy is it", and a header with the word weight gives the search something to match.
Keep units in the header or the cell, never in your head. A column of bare numbers called "Capacity" leaves the bot to guess whether 40 means litres, kilograms or watts. It will often repeat the number without a unit, which is worse than saying nothing.
Use words for yes and no. "Yes" and "No" survive the trip. Ticks, crosses, colour fills and conditional formatting do not, because a CSV holds text only.
Keep cells short and factual. Long marketing descriptions make rows long, and a long row can be split across two passages so that the product name and the fact the shopper wanted land in different places. If you need descriptive copy, keep it on the product page and let the crawler read it there.
Cleaning the export
Spreadsheets that look fine on screen often export badly. Check these before uploading.
- Remove extra header rows, titles and notes. The first row must be the column headers. A sheet that starts with "Spring range 2026, prices ex VAT" will use that line as its headers.
- Unmerge merged cells. Merged cells export as one value and a row of blanks.
- No line breaks inside cells. The file is read line by line, so a cell with a line break in it splits the row in two and scrambles the columns after it.
- Commas as the separator. Some regional spreadsheet settings save CSVs with semicolons. The upload reads commas, so a semicolon file arrives as a single column. Values that contain commas are fine as long as they are wrapped in quotes, which spreadsheet programs do for you.
- Save as UTF-8. Accented letters, the degree sign and currency symbols come through correctly when the file is UTF-8 encoded.
- Avoid exactly two columns. As described above, a two-column file is read as questions and answers. A sheet with only Model and Price columns should gain a third column, even a simple Category.
What a spreadsheet will not do
Two limits are worth planning around.
Comparisons across many rows. "Which of your tents is the lightest?" needs the bot to see every tent's weight at once. It only sees the five closest passages, so with a hundred rows it may compare five of them and give a confident but incomplete answer. For the comparison questions you hear often, add a Q&A pair: "Which is your lightest tent? The Trail 40 at 1.2 kg." The guide to building an FAQ that doubles as chatbot training data explains how to write those well.
Anything that changes daily. An uploaded file is a snapshot. Stock levels, delivery dates and flash prices go stale, and the chatbot has no live connection to your stock system. Leave those columns out, and let the bot point shoppers to the product page or your team for availability.
Updating the file
When specs change, delete the old file's entries from the knowledge list in your dashboard, then upload the new version. Each upload adds its rows as a new source named after the file, and uploading a second copy does not replace the first. If both stay, the bot may quote either.
After each upload, ask five or six real questions: one about a single spec, one about a unit, one comparison, and one about a product that is not in the sheet. The last one should produce a polite "I don't have that information" rather than a guess. Questions the bot struggles with are logged as knowledge gaps, which tells you which column or Q&A pair is missing.
Spreadsheet, pages or both
A spreadsheet is the best source for structured facts: dimensions, materials, compatibility, capacities. Product pages are better for the descriptive answers: what it is for, who it suits, how it feels to use. Most stores do best with both. Crawl the core pages, upload the spec CSV, and add Q&A pairs for comparisons. The post on training a store assistant on your product catalogue covers the wider catalogue approach, and the docs on training your bot list every upload option.
Try it with your own products
Start by seeing how the bot handles your existing pages. Paste your address into the free trial builder, which reads up to ten public pages and gives you a working assistant to question. Then, once you have an account, upload a cleaned spec CSV and ask the same questions again to see what the structured data adds.