Yes or No Questions AI Can Answer About a Website

A good yes or no header asks about one thing, names what to look for on the website, and says what to answer when the site is silent. Write it that way and you can filter 5,000 rows by the answer, then read only the rows the AI was unsure about.

Updated October 10, 2026.

By the end you will have a header template to paste, a way to pick your review line from a 30 row sample, and a spot check that puts a ceiling on how wrong your "trusted" rows could be.

Ask about one visible thing, not about "good"

"Is this a good lead?" fails because the AI has to guess what "good" means to you. "Does this company sell software to other businesses?" works because the answer sits on the website.

Use this template:

Does [company] [one visible thing] on its website? If [edge case], answer [default].

Start with "Does", "Is" or "Has". Name something a person could point at: a pricing page, a "book a demo" button, a careers page, a customer login.

Example: the headers and confidence numbers below are illustrations, not results from a real list.

Here is how the third header might come back:

Read the 0.54 row by hand. The other two you can leave alone for now.

General spreadsheet guides give the same starting advice. A guide on adding yes or no answers with ChatGPT says "Is task X completed based on the data available?" beats "Should I do this task?". The Grist AI assistant docs say a clear column header like "Net Profit" gets better results than a label like "D", and that a True or False result calls for a yes or no question. That covers the basics. Company websites need three more rules: split columns, set a silent-site default, and test before the full run.

Split any header that holds two ideas

"Is it B2B and based in the US?" can come back Yes when only half is true, and you can't tell which half. You end up rechecking the list by hand.

Splitting costs more row-questions. One row-question is one question against one row, so the cost is rows times columns. A 5,000 row list with 3 yes or no columns is 5,000 x 3 = 15,000 row-questions. That is cheaper than the hand checking a muddy header causes. To turn the finished columns into a filter, see how to filter a lead list by ICP fit.

Decide what the AI says when the site is silent

A plain Yes or No forces a guess when the site doesn't say. You can't tell the guesses from the solid answers. Cleanlab's guide to yes or no decisions puts the problem this way: with only a Yes or No output, it is hard to control false positive and false negative rates, and a confidence score lets you pick a threshold that gives the error rates you want.

Pick the silent-site default by which mistake costs you more:

• A wrong Yes costs more (a bad lead gets your sales email): default to No, then review the low confidence Yes rows.
• A wrong No costs more (you would miss a good lead): default to Yes, or add a separate "unclear" column, then review the low confidence No rows.
• You want three honest outcomes: write the third into the header.

Example header with a third outcome:

Does the site name a price or a pricing page? Answer Yes, No, or Not stated if the site has no pricing information.

If you run several yes or no columns, score the answers in a spare column so you can rank rows. Research on prompting models for Yes, No or Unknown answers gives them numeric values of 1, 0.5 and 0. Do the same: Yes = 1, Not stated = 0.5, No = 0.

Example: three columns scored this way give every row a score from 0 to 3. A row that reads Yes, Yes, Not stated scores 2.5. A row with 3 is your strongest shortlist candidate.

Test 20 to 30 rows before you run the whole list

Every wasted full run burns row-questions. A 500 row sample with one column is 500 row-questions, which is the size of the free plan. A 160 row sample with 3 columns fits too.

1. Pick 20 to 30 rows.
2. Run the header.
3. Read each answer against the site.
4. For every row where you disagree, write down the word in the header that caused it.
5. Reword that word and run the same rows again.

If many rows come back unsure on one header, suspect the wording before the list. Fix the header first. Writing headers as plain-English questions means there are no prompts to debug, only the sentence in the column name.

Draw your review line from the sample

Don't guess a confidence cut-off. Sort your 30 sample rows by confidence and read them from the bottom up. Note the confidence level where you stop disagreeing with the answers. That number is your review line for the full list.

This is a working rule, not a published standard. Set the line per column. A clear question like "Is there a careers page?" can be trusted lower down the scale than a fuzzy one, so one cut-off for every column will waste your time on some and let errors through on others.

A Yes at 0.95 and a Yes at 0.55 are different bets. Merge them and you either email bad leads or check everything. What to do with the rows below your line is covered in how to know which AI answers are wrong.

Spot check the trusted rows with the rule of three

Rows above your review line are the ones you won't read. A random check of some of them tells you how bad the pile could be.

The rule of three: if you check n rows and find zero wrong answers, the 95 percent upper bound on the error rate is 3 divided by n. Wikipedia's page on the rule of three describes it as the interval from 0 to 3/n. Check 100, find none wrong, and the rate is probably under 3 percent.

To pick how many rows to check, use n = 3 / target rate:

Example: say 4,000 rows sit above your review line for one column. You check 100 at random and find none wrong. The ceiling is 4,000 x 0.03 = 120 rows that could still be wrong. That is a worst case at 95 percent confidence, not a forecast and not a promise that the rest are right.

Steps

1. Filter to the rows above your review line for one column.
2. Add a helper column with =RAND(), sort by it, and take the top n rows.
3. Open each site and write Right or Wrong next to the answer.
4. Count the Wrong marks. Zero: divide 3 by n and write it down. One or more: reword the header or raise the review line, then sample again.
5. Keep a note: "Header: [text]. Checked [n] high confidence rows on [date]. Wrong: [count]. Ceiling: [3/n]."

Where this breaks

• Picking rows by eye, or the first rows of the file. They aren't random, so the ceiling no longer holds.
• Checking 10 or 20 rows and quoting 3 / n. The same Wikipedia page says the approximation is good when n is greater than 30. A clean small sample can look safer than it is. Take 30 at least.
• Reusing a clean sample after rewording the header. The old check belongs to the old question.
• Sampling the whole list. You spend checks on rows you would have reviewed anyway.

Checking 100 rows takes about 100 minutes at one minute a row. Swap in your own pace. When the list is ready, run the full list.

Do this in the next ten minutes

• Pick the one yes or no question that decides who gets contacted.
• Write it with the template: one visible thing, one edge case, one default.
• Choose the default by asking which wrong answer costs you more.
• Pull 30 rows from your list into a sample file.

That is enough to run your first test.

Test the header on your own rows

Columns works this way: you type the question as a column header, and every answer carries a confidence number, so your review line comes from your own data. The free plan includes 500 row-questions with no card, and cached answers are not billed again when you rerun a refined list. Start free, no card and put your header on a sample of your own CSV.