This post does two things. It names a technique for getting a company cited by AI answer engines, and it puts a date on the idea so that when someone else describes it next spring, the record shows where it came from. I am writing it before we have proven the technique works. That is deliberate, and I will explain why at the end.
The problem it is aimed at
Since June, we have been measuring, every week, whether ChatGPT, Claude, Gemini, and Perplexity name Coworkers.Global, when someone asks the eight questions a buyer of ours would ask. We run each question through each engine''s model API with live search on, discard any answer where the engine chose not to search, and count. As of September 13, 2026, the count is thirteen straight weeks at zero named mentions on that eight-question set (0 of 24 valid answers in the latest reading).
We are not unusual. The industry''s favorite structural fix, the llms.txt file, was measured across 137,000 domains in May 2026: 97% of the valid files received zero requests from AI retrieval bots that month. Our own site has one. It is well formed. It has not moved the number.
What the measurement did show, once we pulled the raw answers apart, is that the engines were not ignoring the category. They were answering the questions from other people''s pages, mostly roundups and case studies on aggregator domains we have no presence on. The engines had a source for every question. It was never us.
Nobody was home at our address.
What a FAQtopus™ is
One HTML page. On it, the eight buyer questions, verbatim, each as a heading, each followed by a short answer that could stand alone if an engine lifted it out. Under each short answer, a link to a full post that answers the same question at length, with the sources shown. On each of those posts, a small block that gives a reader the one fact, the date, and a pre-formatted line for citing it.
A head and eight arms. Hence the name. The domain, faqtopus.com, is live as of today.
That is the whole spec, and on its own it is not new. Answer-first pages are the oldest advice in this field, and the evidence for them is the strongest evidence there is: engines lift blocks that look like answers. What I think is new is the constraint that the questions on the page are the same eight questions we measure against, string for string, key for key. That one rule turns a content page into an instrument. When we re-measure in two weeks, every arm reports separately, and we can see which question moved and which did not.
Nothing else we have shipped has that property. An llms.txt file cannot tell you which of its lines an engine read. A schema block cannot tell you which field mattered. Eight questions with eight keys can.
The build, in four steps
- Develop the eight questions, plus a ninth that never appears. The eight are the same locked buyer questions we already run every week through four engines, one per category we sell into. The ninth exists only to be held back, measured the same way, and never given an arm or a place on the page.
- Ascertain demand and phrasing volume. Before any of the eight gets an arm, check that the exact wording is actually asked, in Google and in AI prompts both, not assumed from a changelog written weeks earlier. This step already caught one bad question: the wording we had for the grant-writing question pulled next to no search volume and was fading. The real term underneath it, once we checked, pulls a real and growing number. We reworded the question instead of writing a post nobody was asking for.
- Draft all eight full blog posts, each with citations and a "cite this" block. Every figure sourced. No post ships answering a question we have not separately confirmed a buyer actually asks.
- Assemble the FAQ page: eight questions, eight summarized answers, eight links out to the full posts. The head is the last thing built, not the first, because it is only as good as the arms underneath it.
The arm we cut
The original spec, as it came out of a conversation this morning, ended with "an encouragement to create backlinks." I removed that line, and I want to be specific about why, because it is the kind of edit that looks like pedantry but isn''t.
A sentence asking readers to link to you converts at roughly zero, and both human readers and retrieval systems have learned to read it as a signal that the page needs the links more than it deserves them. What earns a link is something a reader wants to quote: a dated figure, a number nobody else has published, a definition that settles an argument. So each arm ends with a "cite this" block instead of an ask. Here is the fact, here is when we measured it, here is how to cite it. If the fact is worth citing, the link follows. If it is not, no amount of asking would have helped.
That is my view, not a finding. The finding comes later.
How we will know if it works
Here is the test, written down before the page exists so we cannot move the goalposts afterward.
Nine questions are locked and keyed, not eight. Eight get a place on the page and an arm behind them. The ninth gets neither. We measure it on the same schedule, with the same engines, and never show it to a reader or a crawler anywhere. We have a baseline for all nine: three runs per question per engine, taken the week before the page goes live. We publish the head and the eight arms. We re-measure at fourteen days and thirty days, same nine questions, same engines, same three runs. If the eight on-page questions move outside the run-to-run variance we saw at baseline and the ninth does not, the technique did something. If all nine move together, the engines changed underneath us and the page gets no credit. If nothing moves, I will say so in the follow-up post, with the numbers.
The first host is our own site. Thirteen weeks at zero is a clean starting line, and it doesn''t waste a customer''s patience on an unproven idea.
Current view, subject to change
My base case is that the head will get cited before the arms do, and by Perplexity first, because that engine retrieves and re-ranks live pages and rewards answer-shaped blocks. I expect Claude to be the engine that quotes the arms, because it favors depth and third-party corroboration, and the arms are where the sources live. I doubt ChatGPT moves at all in the first thirty days; its citation rate is low, and Wikipedia and a handful of large publishers dominate its source mix.
What would change my mind: a control question that moves as much as the treated ones. That would mean we were measuring the weather, not the technique. I would also revise the whole idea if the head gets cited, but the arms never do, because then the "cite this" block is decoration and the technique is just a good FAQ page with a mascot.
Final thoughts
Most of what is written about AI search optimization cannot be checked. A firm publishes a tactic, a screenshot, and a testimonial, and no one can say afterward whether the tactic did anything, because nothing was measured before, nothing was held constant, and nothing was held out. I would rather publish an idea with a test attached and risk reporting that it failed. The record of a failed test is worth more to the next person than the record of an unmeasured success.
So: one FAQ, eight arms, eight keys, one control. Date of record: September 14, 2026. The numbers follow in October. Check it out faqtopus.com.
Regards,
Charles Stack
Cite this: "The FAQtopus™ technique (one FAQ page keyed to a measured buyer-question set, one sourced post per question, one held-out control) was first described by Charles Stack, Coworkers.Global, on September 14, 2026." Source: this post. Measurement basis: 13 weekly readings, June to September 2026, four engines via model API, no-search answers excluded.
"FAQtopus" is a trademark of Coworkers Global, Inc.