Published August 25, 2026Updated August 28, 2026
August on August:
Measuring Our Own AI Search Visibility
We sell AI search visibility, and in August 2026 our own site had none. This is the log: what the baseline actually measured, what we changed, and what has and has not moved since. It is updated as it happens, zeros included, and every figure carries the date and the method that produced it.
01 — The baseline
Eight queries, zero appearances
On August 11, 2026 we ran the first measurement against our own domain: eight queries, phrased the way a buyer would phrase them, checked directly in live search results. The domain appeared in none of them. Searching the exact domain name returned French villages beginning with "Beau".
The site behind that result had one indexable URL. Its entire body text was around 460 words — a single page carrying five service lines, roughly twenty words each. There was nothing for a search engine to rank beyond the home page and nothing for an answer engine to retrieve at all, because a passage that resolves a question on its own cannot be written in twenty words.
Three of the findings were not gaps but failures. An edge function was answering visitors from Spain and Argentina with a 404 and a noindex header — two of the three offices the site itself lists on the home page. The Open Graph image and the favicon both returned 404. And the structured data listed another agency's Instagram account as ours, which is the single worst thing a small brand can tell an engine that is trying to work out which entity it is looking at.
02 — What we changed
From one indexable URL to eighteen
Everything below was published and then verified against production — not against the repository. A correction that only exists in a commit has not happened.
Aug 11–19 — The failures first
The noindex on the geo-block was removed for crawlers while the commercial block stayed in place, and the two 404s were fixed. Verified live: a crawler with a simulated Spanish or Argentine origin now receives a 200 with no blocking header.
Aug 15–17 — Three service pages
Three service pages went up on their own URLs — generative engine optimization, custom AI agents, paid media — between August 15 and 17, each with structured data and its own set of questions. Measured on August 25: 1,421, 1,531 and 1,651 visible words.
Aug 23 — An industry page
On August 23 an industry page went up for pharma and healthcare, the vertical where six of our twenty brands already are and where no agency in the result set shows the same profile. Measured in production: 1,721 visible words.
Aug 24 — The first article
On August 24 the site published its first article, after a measurement showed that the sources these engines cite are articles rather than service pages. Measured on August 25: 1,909 visible words. The piece you are reading is the second.
A crawl run on August 25, 2026 with a Googlebot user agent confirms the state: nine HTML URLs, all returning 200, each with a title, a meta description, a canonical URL, exactly one H1, no blocking robots directive, and every structured-data block parsing. The site that measured zero on August 11 no longer exists in that form.
03 — What has not moved
Twenty-one rounds, still zero
Twenty-one measurement rounds on twenty-one dates between August 11 and September 2, 2026 — the baseline, and then one a day from August 14 on. The domain did not appear in a single tracked query on a single date. Ten questions put to answer engines over the same period: not named in any answer, not cited as a source in any answer.
That is the honest headline and we are not going to dress it up. What the twenty-one rounds did produce is the thing a single run cannot: the result sets are stable. The same handful of domains occupies the two commercial queries every day, which means the variable we were changing over those weeks — pages on our own domain — is not the variable that controls those results.
The second finding is about the shape of the results rather than our position in them. In the transactional query, most of what appears is roundups and directories — "top 10", "12 best" — rather than agency pages. On August 25 we asked an answer engine almost the exact title of the article we had published the day before, and all six results were articles on companies' own publications. Both of those are lists you enter by being written about somewhere that is not your own domain.
The brand query points the same way. Asked for our name, the engine returns ten unrelated companies called August and states outright that it has no information on the one we are asking about. It understands the question. The entity is simply not established anywhere it can read — and no amount of editing our own pages establishes it, which is the constraint we are working against now.
One measurement was retired rather than reported. Two of the five witness queries returned the same result set for five consecutive days, which means they are one measurement counted twice, not two data points. It was replaced at the end of the month. A tracking list that quietly double-counts is how a report ends up looking more solid than the evidence under it.
| Tracked query | beaugust.co | What comes back instead |
|---|---|---|
| AI marketing agency Miami | Absent, 17 in a row | The same seven every day: Relevance, UFO, Digital Agency Network, Miss Ink, Up Digital, HarmonyTec, iAnalyst. |
| marketing agency Miami AI | Absent, 15 in a row, then retired | Exactly the same seven results as the query above, in the same order, for six days running. That is what told us this query was not measuring anything the first one wasn't — it has since been replaced. |
| generative engine optimization agency | Absent, 17 in a row | Thrive, Directive, Digital Elevator, GreenBanana, Minuttia, Percepture. Four of the six are listicles or directories. |
| agencia de marketing con IA | Absent, 17 in a row | A Spanish-language result set with no source in common with the English one: Single Grain, Vilma.AI, Media Power, Screen Media Group, DigitaltoGrow. |
| August marketing agency beaugust | Absent, 17 in a row | Ten unrelated "August" companies — August Marketing Studio, August Strategy, August United, August Australia, whoareaugust — and an explicit note that it found nothing for "beaugust". |
These are manual search verifications, personalised and localised. They are not ranking positions and should never be filed as such. Source: the change log kept for beaugust.co, one entry per day.
04 — Search Console
Two days of data, and one page Google will not fetch
The property was verified by DNS from the start, but the account was only opened on August 26, 2026. That is the day this page gained a second source — and the first thing that source showed is how little a new property has to say.
The performance window is two days long: August 25 to 26. That is all Google backfilled on verification, and with the report's usual two-day lag it is everything available on August 28. In those two days, one impression, no clicks, and an average position of 5. The impression went to the home page. Which query produced it cannot be known — Google anonymises very low-volume queries, and the queries table reads no data.
A position of 5 over a single impression is the position of one event, not an average with anything behind it, and it is not a ranking baseline. It is written here at that size on purpose. The temptation with a new property is to report the first flattering figure it produces; the honest reading is that two days cannot be compared with a full month of anything, and that comparison is exactly what the next report will be tempted to make.
Indexing is the part that did move. Inspected URL by URL on August 28: eight of the thirteen submitted URLs are indexed. Two days earlier it was four of seven inspected, with three sitting in Discovered — currently not indexed and no crawl date at all. Four pages came unstuck in those forty-eight hours, and two of them had been published on August 26 and were indexed in under two days.
One URL will not move, and it is the one that sells the main service: the generative engine optimization page. Thirteen days after publication and two days after an explicit indexing request, its last crawl date is still blank. Google has never downloaded it. The live test on that exact URL returns that the URL is available to Google and that the page can be indexed, with a valid breadcrumb — so there is no technical impediment: Google can fetch it right now and is choosing not to in its regular pipeline. It also carries more internal inbound links than any other page on this site, nine of them, which rules out internal linking as the cause.
That single blank crawl date states the constraint more plainly than anything else on this page. A page can be technically perfect, well linked and explicitly submitted, and still not be worth fetching. What decides that is not on our domain — which is the same conclusion the twenty-one zero readings above arrive at from the other direction.
Core Web Vitals still report no field data. The property is new and the report needs real traffic before it can say anything — which is an answer in itself, and the reason there is no speed figure anywhere on this page.
05 — Method
Where every number here comes from
The method is the part of this that transfers to a client account, so it is worth stating plainly.
Two sources, kept apart, both dated. The first is live verification of search results and crawls of our own domain, and it is where every figure before section 04 comes from. The second is Search Console, opened on August 26, 2026 and reported on its own in section 04 — because its window is two days long and covers almost none of the period the rest of this page describes. What an answer engine says is recorded apart from both, because they are different measurements of different systems and averaging them produces a report nobody can act on.
The zeros get recorded with the same care as anything else. Twenty-one identical results are not a wasted three weeks — they are the evidence that the answer is stable and that the thing we changed was not the thing that mattered. Deleting them because they look like nothing happened is how a team ends up doing the same fix three times.
This page updates as the measurements come in. If the approach turns out to be wrong, that will appear here too, on the same day it becomes clear.
06 — Sources
What we relied on
Every figure above is our own measurement on our own domain, dated. These are the outside references behind the method.
- Don't Measure Once: Measuring Visibility in AI Search (GEO) Why a single reading of an AI answer tells you almost nothing, and repeated readings on a fixed cadence tell you a lot. It is the argument for keeping this log at all.
- Aggarwal et al. — GEO: Generative Engine Optimization KDD 2024. The paper that named the field.
- Google — Page Indexing report: status reasons Google's own definition of "Discovered — currently not indexed". Three of our URLs carried it on August 26; one still does on August 28, with no crawl date at all.
- Google Search Central — Crawl budget management How Google decides how much crawling a site is worth. It is the mechanism behind a page sitting thirteen days without ever being fetched.
07 — Questions
Frequently asked
Why publish a case study where all the numbers are zero?
Because the zeros are the part that is normally hidden, and hiding them is what makes this category hard to buy. Every agency publishes the month the graph turned up and none publishes the weeks before it. We are running the same service on our own domain, starting from an audited zero, and posting each measurement whether it moved or not. If the approach does not work, that will be visible here too, which is the only thing that makes the rest of it worth reading.
Are these Search Console positions?
No, and the distinction matters enough that we label every number with it. What is recorded here is live search-result verification: we run the query, read the results, and note whether the domain appears. Search Console reports impressions and clicks from Google's own logs, which is a different measurement of a different thing. The account was opened on August 26, 2026, and its first window is two days long — Google backfilled no further on verification — so it reports one impression and no clicks. Those figures are in section 04, kept apart from the verification series above, because averaging two sources produces a report nobody can act on.
Twenty-one rounds without movement — is that a failed project?
It is a result, and reading it as a failure is the mistake that leads to repeating the same fix. Twenty-one identical measurements say the result set is stable, which means the variable we were changing during those days is not the variable that controls it. That is worth more than a single noisy improvement, because it rules something out. Two weeks is also short: pages published on the 23rd and 24th have not been crawled, indexed and evaluated yet, and nothing that depends on other people mentioning us could have happened in that window at all.
What actually changed on the site in those weeks?
The site went from one indexable URL to eighteen, and three of the corrections were failures rather than additions: an edge function was returning a 404 with a noindex header to visitors in Spain and Argentina — two of the three offices the site itself lists — the Open Graph image and the favicon both returned 404, and the structured data pointed at another agency's Instagram account as if it were ours. Those three were fixed and verified in production. The remaining work was writing: fourteen new pages, the first five of them between roughly 1,400 and 1,900 visible words each.
Why doesn't publishing more pages move the result by itself?
Because for the queries we are tracking, the pages that occupy the results are not competing on the same axis. In the two commercial queries the incumbents are the same handful of domains every single day, and in the transactional one most of the results are roundups and directories rather than agency pages — a list you enter by being mentioned in someone else's article, not by writing a better page of your own. Publishing is necessary: there has to be something to retrieve. It is not sufficient, and our own log is the evidence.
Would you run a client account the same way?
It is the same routine, which is why we run it here first. A fixed battery of queries phrased the way a buyer phrases them, measured on a fixed cadence, every result recorded with its date and its source — including the ones that did not move. The difference is only what the client sees: the same log you are reading, on their domain, with the reasoning for each change written down next to the change. If a number in a report has no source and no date attached, it should not be in the report.
Find out what the engines say about you
The audit we would run on your domain is the work described above, on your questions: the baseline measured and dated, the failures separated from the gaps, and every result recorded whether it moved or not.