Does improving AI visibility work? Measuring from a United States baseline of zero

Yes, though in the United States the thing worth measuring is usually not a rate. A run on 15 October 2026 asked 102 buying questions in each of three markets, and in the American half of it neither engine used our own domain as a source even once. A wider set followed on 16 October 2026, 210 questions per market, and ChatGPT returned our domain as a source once in its 210 American readings while Gemini did so in none of its 210. A great many American businesses are starting from exactly that position, and a percentage change cannot be computed from zero. So the first useful output of an American before and after test is a date and a URL: the day a named page of yours first appeared as a source for a named question. Rates come later, once there are enough citations for a rate to mean anything.

What zero and one look like, with the dates attached

597 usable readings of 612 attempted, from 102 questions in each of three markets, make up the earlier run, published in what AI assistants actually cite across 597 readings of ChatGPT and Gemini in three markets. Our domain was cited 3 times in those 597 readings and none of the 3 were American. The second run, on 16 October 2026, widened the question set to 210 per market in 21 batches, 630 in total, and produced 1,260 readings with none discarded. Our domain was cited in 6 of them, which is 0.5%, and 1 of those 6 was American.

Reading	Date	US detail
First run	15 October 2026	102 questions per market. Our domain cited zero times in the US, on ChatGPT and on Gemini.
Second run	16 October 2026	210 US questions per engine. One ChatGPT reading used our domain as a source; no Gemini reading did.
Both runs, Gemini	15 and 16 October 2026	0 citations of our domain in 921 Gemini readings across the two runs.

Two question sets built independently, the second one designed to go at the first one's blind spots, arriving at the same answer about the American market. Agreement reached that way is worth more than either run alone. The American cut of the first run, with the full source mix behind it, is published in what the two engines cited for United States buyers.

Why a rate is the wrong first measure at this end of the scale

Two hundred and ten readings put a single answer at about half a point of the total. So a company going from zero citations to one has a percentage that looks like noise and an event that is entirely real: a specific page of theirs became a source for a specific question on a specific day. The percentage understates it and a percentage change cannot even be calculated, because the starting figure is zero. Report the event.

That gives an American test three outputs, and they arrive in this order.

• A first citation date. The day a named URL of yours first appeared as a source for one of your frozen questions, with the engine named and the question quoted. This is the first thing that happens and it is binary, so no sample size argument can be had about it.
• A count of distinct questions you appear for. Two questions out of forty is a count anyone can check. It becomes meaningful well before a rate does.
• A rate with both fractions printed. Worth reporting once the count is large enough that one reading moving does not move the headline. Until then, publish the count and the denominator and leave the percentage out.

What holds the American slots, and what that means for the test design

The reason an American baseline so often reads zero is visible in the source mix. Of the 20,143 citations recorded on 16 October 2026, across 4,947 distinct domains, the single busiest was developers.google.com, present in 446 of the 1,260 readings. Behind it, support.google.com at 328 and help.openai.com at 269, both a platform writing about itself. Those are platforms answering questions about themselves. Any frozen question asking how an engine works has a winner already, and a dozen of them in an American test will flatten the result however good the work is.

Underneath that layer the picture is different: 82.0% of the same 20,143 citations went to a vendor or agency own site, and review directories took 0.7%. The American slots that can change are held by self published pages, which is the route that is open. Sort your frozen questions into the ones a platform will always answer and the ones a supplier can answer, measure both, and judge the work on the second group.

What to freeze before the first American measurement

• The exact wording of every question. Not the topic. A question retyped differently at the follow up is a new question and its result is not comparable.
• The country, inside the question rather than in a setting. Country settings are unreliable here: one can accept the setting, report success and answer about another market entirely. Put the United States in the sentence, then read the answer for dollars before counting it.
• No brand anywhere in the question. Supplying the name tests recall of something you just handed over. Phrasing it the way an American buyer phrases it tests whether anybody gets put forward, and who. Keep that version.
• The engines, named and reported apart. The same American questions gave 1 citation on ChatGPT and 0 on Gemini on the same day, and one blended American score would have described neither.
• The raw answers, saved. 78 of the 102 build spec conversations we ran asked for the prompt set to be published so the work can be re-run, and 68 asked for failures and negative results to be shown. Saving the answers is what makes both possible later.

Settle the measurement in the contract, before anyone writes a page, so the definition of working is not negotiated after the numbers arrive. What belongs in that paperwork, and the questions that make two American proposals comparable, is covered in running a search procurement you can defend.

What these two runs establish, and what comes next

These two runs establish a baseline and its stability. No intervention sat between them, so what moved from zero to one is the baseline itself, read twice. The American figure held when the question set was rebuilt and doubled, and that steadiness is the thing a before and after test needs before it can read anything at all: a baseline that wanders tells you nothing about a change made on top of it. The intervention study is the next measurement, a baseline rate and a day 90 rate either side of a dated change, and it will carry its own numerator and denominator. Everything here belongs to ChatGPT and to Gemini: Google AI Overviews, Perplexity and Claude never saw these questions, so no line here reads on them, and the five surfaces we cover for customers stay separate from these two. No American figure by trade exists in the data either, because the questions aimed at particular trades were worded generically.

Questions about measuring US AI visibility from a baseline of zero

How long until results appear?

Agree the follow up date before the first reading is taken, and print it next to the figure. What to watch for is not a moving average but a first citation date on one named question, because that is the event that actually happens first.

Can a content change produce a measurable improvement?

Of the 20,143 citations that run recorded, 82.0% landed on a page somebody had published about themselves. Publishing is the route. Which page moved which American question is what a frozen set exists to answer, and only your own set can answer it.

What if the American count goes back to zero?

Then you learned it early and cheaply, and it goes out as a rise would have. Our own first run returned 3 citations in 597 readings with none of them American, and it is published that way, because a programme that only shows its improvements is a programme nobody can check.

How do I separate our work from the engine changing on its own?

Name the questions nobody is allowed to touch before the work starts, and measure them on both dates alongside the rest. They are the control, and in a month when an engine shifts its own behaviour the control is the only thing that will show you.

Is one citation in 210 readings worth reporting to a board?

Report it as one reading out of 210, with the engine named, the question quoted and the date attached. Stated that way it is a checkable fact rather than a percentage taken on trust, and a number arriving with no denominator is the thing a board should push back on.

What the first American citation actually tells you

The day a page of yours becomes the source an engine uses for an American buying question, you learn something a rate never tells you: which page, and for which question. That pairing is the instruction for the next quarter of work, and it is why the first citation date is worth more at this stage than any average. Measuring that, and writing the pages that produce it, is what we are paid for. The numbers above are our own: one citation in 210 American ChatGPT readings on 16 October 2026, and none in the American readings of the run before. Every figure above has its denominator and its day beside it, and the next round asks the same questions again.