For about twenty years, search marketing ran on a number anyone could read. You could find a page at position seven, rewrite its title tag, and a month later find it at position three, with a green line in the rank tracker to celebrate. Then OpenAI opened ChatGPT to the public in November 2022. Perplexity built its product around answers with footnotes, and Google put AI Overviews above its own results in 2024. None of those answers came with a position one. They named a few sources and moved on.
So the rank tracker lost its unit. An AI answer either cites your page for a given question or cites someone else, and there is no seventh place to climb out of. Somebody will sell a visibility score out of 100 for this, and you will be invited to buy it. We think the useful number is smaller and much duller: how many of the questions buyers ask end with your page in the citations. It moves one question at a time. A spreadsheet can hold it.
The question list starts on sales calls
You heard the best list of questions out loud on your sales calls long before anyone typed it into a keyword tool. On those calls, your buyers asked what it cost, how it compared with the tool they already paid for, how long setup took and what happened to their data if they left. Those are the questions they later put to ChatGPT in roughly the same words. Keyword tools return the same questions flattened into two-word phrases nobody says.
Write down fifteen or twenty of them, in the buyer's phrasing. Then ask each one in ChatGPT, Perplexity, Gemini and Google's AI Overviews once a week, by hand. Doing it by hand sounds quaint. But ChatGPT, Perplexity and Gemini change their answers with the wording, with the week and with whatever they retrieved that day, and you only catch the odd sentence about your company by reading them. Keep the conditions steady from week to week, with the same wording, the same logged-in or logged-out state and the same country where possible.
One row per question
The log is a plain sheet with one row per question, per assistant, per week. In each row, you note the date, the assistant, the sites the answer cited, whether yours was among them and any sentence it printed about the company. With four assistants and twenty questions, you fill eighty rows a week, which sounds like a lot until the first row turns up something wrong.
The cited sites column grows into a list of who wins the questions that matter, and it often names rivals a founder never thought of as rivals. That list alone was worth the hour for anyone who has kept one.
Of the five columns, the last one carries the most weight. A citation that gets the pricing wrong is worse than no citation at all, and a tally of yes and no would have counted it as a win.
Analytics already caught some of the clicks
Your analytics held some of this before you built a sheet. Check what OpenAI, Perplexity and Google publish about referrers before you trust the split between referral and direct, since those pages change now and then. Treat the referral number as a floor. It counts the people who clicked, and an answer that cites a page can do its work without anyone clicking.
How a page is written moves the citation
Rank alone does not settle who gets cited. In 2026, Rustem Kakimov, Xing Tan, Jonathan Gillham and Narcis Bejtic audited citation behavior in Google AI Overviews on "Your Money or Your Life" queries, the health and money questions Google treats with extra care, drawn from the MS MARCO Web Search dataset. They tracked where each document ranked in retrieval and whether the summary cited it.
That is one engine, one family of high-stakes queries and an observational study. It measured where a document came from, a machine or a person, and it did not test answer-first paragraphs or any other style rule. What it supports is narrower and still useful. We think part of that something is shape, since machine-written pages tend to open with a tidy answer, though Kakimov and his coauthors never tested that and we would not stake a quarter on our reading of it. It is also a funny result for everyone who predicted Google would bury anything a model wrote. We'd rather not read it as a reason to publish more of that.
Put the answer in the first paragraph
The part a founder controls is the opening. Our earlier post, "Write the first paragraph so ChatGPT can quote it," argued that the first paragraph should hold the answer itself, with a name, a number and where the number came from. We took a line about our own way of writing and wrote it both ways.
Before. We think hard about how we write at Propaganda, and we have tried to make every sentence count.
After. Propaganda checks four assistants each week, as the log above shows.
We wrote the first version politely and gave ChatGPT nothing to quote. The second put a name, a number and a source into two sentences. The second version holds a passage that ChatGPT or Perplexity could lift whole for anyone asking how Propaganda works.
Fix the page the answer read
When you find an answer that cites you and gets the company wrong, start with the page it cited. Open it. The wrong price or the retired feature may be sitting right there, in a post written before the last launch, and the assistant repeated it faithfully. Fix the page, note the date in the sheet and ask the same question the following week in the same words.
If the answer changed, log the new text. If it still repeats the old fact, log that too and keep asking. If you skip the follow-up, you have a guess, and the logged answer from the next week gives you something you know.
The count needs a few pages to mean anything
Below a handful of pages the count is mostly zeros. If your blog has three posts, you will see zero citations for a while, and that zero is a reason to publish before it is a finding about AI search. Our own blog is a few weeks old. When we planned it on 14 October 2026, we aimed for two searches on page one this quarter and expected the other eight to take longer. Our blog started from the same zero in AI answers.
Once you have pages to cite, the sheet is worth its hour. Until then, the hour is better spent writing one.