FotoRegistro already appeared in AI answers in 2026. First place was a toss-up.

first-place appearances in ChatGPT on 20 July 2026

22 of 361361 spontaneous questions identical at both endpoints. It counts when the response recommends the brand in its text and places it first in the list. This endpoint was captured with ChatGPT in anonymous mode.

→

first-place appearances in ChatGPT on 1 September 2026

71 of 361the same 361 questions, 43 days later: 53 appeared and 4 dropped out. This endpoint was captured with logged-in accounts, one of them paid. These are two instrument changes within the window, disclosed in the text, and they prevent attribution of the entire movement to the work.

period

Diagnosis and interview on 1 June 2026. First content batch live on 24 June 2026. The comparison with the same instrument at both endpoints is the 3 July 2026 baseline against the 14 and 15 July 2026 gate: the same questions, judge and contract, with no aggregate movement. Continuous tracking panel since 20 July 2026, with rounds recorded through 7 September 2026.

company size

We did not measure company size. What appears here is operating scale stated by the client in the recorded interview, never measured by us: 40 thousand orders per month, with its own factory in Joinville, Santa Catarina, and sales that are 100% online across Brazil.

universe

395 canonical questions across five areas (institutional and brand, photo books, photo development and printing, framed prints and décor, photo gifts and stationery). The two July campaigns that form the clean comparison measured the SAME 821 unique questions, with different capture counts: 1,759 in the 3 July baseline and 1,649 in the 14 and 15 July gate. The daily panel runs 395 questions across 4 engines, 1,580 captures per round, and the comparable slice between 20 July and 1 September has 361 questions identical at both endpoints.

engines

In the 1 June 2026 diagnosis: ChatGPT and Google AI Overview. In the July baseline and gate, which form the clean comparison: ChatGPT and Claude, captured through the real interface in a clean session, with gpt-4o as judge and the same judge contract at both endpoints. In the daily panel since 20 July 2026: ChatGPT, Claude, Google and Perplexity.

The proposal’s premise was wrong, and the 1 June 2026 diagnosis showed it before the first line was published

FotoRegistro is a Brazilian photo-printing and customisation e-commerce business: photo books, photo development, framed prints and décor, stationery and photo gifts, with its own factory in Joinville, Santa Catarina, and sales that are 100% online across Brazil. What is measured on this page was measured by murmur.marketing, a Brazilian Generative Engine Optimization operation run by Mateus Gomes and hired for this work: it is the same organisation that executes and measures the work, which is why the standard and denominators are written out here.

The proposal that sold this work assumed the brand would be invisible in AI. On 1 June 2026, before writing a line, we checked. We ran buyer questions in ChatGPT with a logged-in profile and in Google AI Overview, with a screenshot of each response. The premise was wrong, and we told the client on the first day.

The brand was cited in 7 of 7 ChatGPT responses that were successfully captured and in 5 of 5 questions where Google AI Overview triggered, out of 6 questions run. In two of each, it already appeared first. Presence was not the problem. The problem was placement and label: AI presented it as the value-for-money option and awarded competitors the quality badge.

The photo-book question, about the product the owners chose as the work’s priority, was repeated three times in a row in the same ChatGPT session. It returned three different leaders: Dreambooks, Nicephotos and FotoRegistro. No brand held first place consistently; it was a toss-up. That was the target, not presence.

This diagnosis is small and worth what it is: one session, few questions, no automatic judge. One initial run came back without the brand and the following three repetitions brought it back, which the report itself records as session noise and prohibits using as proof of absence. Everything else on this page is measured with a judge, a denominator and both endpoints in the same question set.

The category had a dominant voice, and it was not FotoRegistro’s

On 3 July 2026, we fixed a broad snapshot: 821 unique questions, 1,759 captures, ChatGPT and Claude, with each response read by a judge. Across the two engines, the category’s most-cited competitors were Phooto with 551 citations, Nicephotos with 329, Canva with 240, Printi with 213, Dreambooks with 162 and Shutterfly with 104. This list contains competitors only: FotoRegistro’s own score is deliberately left out, because category ranking and client performance are two measurements, and printing both on the same line is how one chooses the flattering slice. Phooto was the dominant voice and the target of every comparison.

The third name on that list is what matters. Canva prints nothing. In a physical-printing category, design software is the third-most-cited competitor, and in the four groups of questions where neither engine recommended the brand, it was the most cited of all, with 75 citations against 39 for second place in that slice. This says what was happening: AI was not choosing a better competitor; it was changing the kind of answer. Someone asking how to make a framed photo receives a tool for assembling the file, not a printer to produce it.

And these were groups the company sells every day: framed photos (27 questions), materials such as acrylic and aluminium (16), stickers (11), and printing from a phone (18). In all four, on 3 July 2026, neither ChatGPT nor Claude recommended the brand even once. Zero. The content plan was designed against these gaps, not against questions it was already winning.

Removing the site’s inflated claim did not move the needle, and one 26-question slice declined

This client’s leading hypothesis was that an inflated claim on the site was hindering citation. On 2 July 2026, we removed it: the overstated production number went, the award badge fixed across the site template went, and article bylines changed from a persona to the organisation. We waited two weeks for AI to recrawl and, on 14 and 15 July, ran exactly the same 821 questions, 1,649 captures, with the same judge and contract, to keep the delta clean.

The aggregate needle did not move. The movement stayed within the noise floor measured days earlier, in the same universe and without changing anything, so it is not a result: it is engine variance. That is the full sentence the data supports, and it is worth more than any percentage that could fit here. There was a cost, too. The group of questions about printing photos online, a core volume line for the brand, declined in both engines in the same direction, in the same 26-question slice. The recorded hypothesis is that the removed authority claim had been helping citation precisely there. We only knew either thing because the noise had been measured first.

The second failure was distribution, and it appeared only because we recounted. On 24 June 2026, we published the first content batch on its own subdomain, and more than a month later the circulating figure said 314 responses cited that content. Recounting by exact host, the real figure was 0 of 1,649 captures: 305 of the 314 were the e-commerce site, not the guide. The cause was proven the same day on three fronts: the e-commerce homepage, across 547 KB of downloaded HTML, did not contain the guide address once; 356 of the 357 articles had no body link; and Search Console had never been verified.

Citation arrived late, engine by engine, and the crawler that read the most cited the least

The guide’s first citation appeared in Perplexity on 22 July, 28 days after the 24 June go-live, and in ChatGPT on 24 July, 30 days later. In the same rounds, Claude and Google remained at zero. Anyone promising a time frame for AI to cite new content is guessing: here it took nearly thirty days, and happened in two of the four engines.

Arrival order did not follow who read. In the 24 and 25 June access log, after separating a scanner that rotated fake user agents and inflated the count, legitimate AI crawling was ClaudeBot with 79 requests, 24 to articles, covering 20 distinct URLs, against 3 OpenAI requests that did not touch a single article. Claude was the only engine that actually read the content, and it is the engine that stayed at zero for five consecutive campaigns. ChatGPT, whose crawler did not open an article in that window, cited it on 24 July. These are two measurements with different windows, which is why they are in separate sentences: together, they allow only the conclusion that reading and citing are not the same event.

Citation was also concentrated in few pages, and in older pages. There were 38 retrieved URLs, 10.6% of the corpus, and 22 came from the first batch of 24 June. From the newer batch, with 114 articles published on 8 July, only 2. Page age mattered more than volume, changing the order of work: rather than simply publishing more, we began fixing and updating what was already live.

Batch 7 shipped with 25 new articles and 240 articles already live updated with the rechecked Reclame Aqui note, without changing a single URL so that no citation already won would break. Batch 8, between 19 and 21 August, had 52 articles: 46 approved on the first pass, 6 sent for manual review and released later, none rejected. On 2 September, before writing the next batch, we crossed 44 new questions with the archive: 37 of them, 84%, already had a live article answering them, so the plan became 5 new pieces and 16 rewrites rather than 44. The guide had 386 pages live when checked against its own sitemap on 2 September 2026.

First-place appearances in ChatGPT rose from 22 to 71, and the window crosses two instrument changes

On 20 July 2026, we turned on the daily panel: 395 questions, four engines, 1,580 captures per round, every day. Between 20 July and 1 September, across the 361 questions identical at both endpoints, FotoRegistro’s first-place appearances in ChatGPT rose from 22 to 71. Fifty-three appeared and four dropped out, so almost nothing was lost along the way, which is what distinguishes movement from fluctuation.

Now the part almost no one publishes. This 43-day window crosses two changes in ChatGPT’s capture pool, not one. Between 23 and 31 July, accounts changed from anonymous mode to logged-in free accounts, and since 12 August a paid account has run in the pool. The 20 July endpoint was measured with one instrument and the 1 September endpoint with another. Separating the instrument’s effect from the work’s effect would require remeasuring both endpoints with the same pool, and that was not done.

So we cannot say all this movement came from the work, and we will not say it. What we can say is what was measured: the standard, denominator, date, and which part of the instrument changed in the middle. A case that publishes only the upward arrow is hiding precisely this paragraph.