Which agency should you hire
to appear in ChatGPT?

01 — The short answer

The kind that measures, and the kind that promises

If you are choosing a provider to get your business named in ChatGPT and in AI answers, there is one distinction that does most of the work: an agency that treats this as a measurement problem, or an agency that treats it as a promise. The first one starts by reading what the engines say about your category today, writes down the method, and shows you the readings that came back empty. The second one sends a proposal.

That is not a style preference. Nobody — no agency, no tool, no amount of budget — controls whether a specific engine cites a specific page on a specific day. What can be worked on is whether you are citable: whether the thing a customer asks has a plain answer somewhere with your name attached, and whether enough places outside your own site say so. And whether that is improving is a question only a measurement can answer.

So the agency that fits is the one whose first deliverable is a baseline you could repeat yourself. The one that does not fit is the one whose first deliverable is a forecast.

02 — Before you sign

Five questions, and what a good answer sounds like

Ask these in a call, in this order. They take about ten minutes and they are hard to pass without doing the work. None of them requires you to know anything technical.

1. 1. What is my starting point, and how exactly was it measured?

   A good answer names the surface, the method, the location and the date: “thirty reads of the AI Overview block for these ten queries, pulled through a SERP API, geolocated to your city, between these two dates.” A bad answer is a score out of a hundred with no method attached. You are not buying the number — you are buying the ability to repeat the reading next month and see whether it moved. Ask for the baseline before anyone touches your site, because after that nobody can reconstruct it.

2. 2. Which queries will you track, and who chose them?

   The list should be short, written the way a customer actually types, and you should recognise your business in it. If the queries were chosen by the agency alone, they tend to drift toward the ones that are easy to win. Ask for both kinds: the commercial ones that name your city or service, and the plain informational ones people ask before they know who to call. Those behave very differently, and we have published what we saw on the local ones.

3. 3. Will you report the readings that come back at zero?

   This is the question that separates measurement from marketing. An absence is a result: it tells you the engine generated an answer and did not take yours. If zeros are not in the report, the report is a shortlist of the readings that went well. Ours are published — 151 reads across 27 queries between September 25 and October 7, 2026, and beaugust.co appeared in none of them.

4. 4. What part of the work happens off my own site?

   Most of it, usually, and that is the part people skip. Answer engines assemble replies out of what other places say about you: directories, listicles, forums, video, reviews, coverage. On our own record the bottleneck was not missing content — two of the queries we measure are, word for word, the titles of two articles we had already published, and the engine cited neither. Ask specifically which third-party surfaces are in scope, and what gets delivered there.

5. 5. Who owns the accounts, the content and the measurements when we stop working together?

   You should, all three, and it should be in writing before the first invoice. Profiles created under an agency's email, articles published on an agency-controlled domain, and a measurement history that lives in someone else's spreadsheet are three different ways of ending a contract with nothing. The test is simple: ask for the export.

Related: how answer engines choose which sources to cite · what we measured on queries that name a town

03 — The red flag

“We guarantee you'll show up in ChatGPT”

Treat that sentence as disqualifying, and not because it is ambitious. It describes something the seller does not control. The answer you get from a generative engine is assembled at the moment you ask it, out of whatever it retrieves then, and it varies by language, by wording, by location and by day.

We have that variance published with dates. Asked in Spanish on September 13, 2026, one engine answered the hiring question by naming four hireable agencies. Asked the same day in English, it named seven — and not one of the seven appeared in the Spanish list. Zero overlap. The full record, with the six names it returned, is here in Spanish. If a single day and a change of language can replace the entire list, no one can promise you a seat on it.

What can be promised honestly is narrower and more useful: a baseline, a set of queries you agreed on, a cadence of readings, work on specific surfaces, and a report that includes the zeros. If you ask for a guarantee anyway, the most trustworthy answer you can get is someone explaining why they will not give you one.

There is a second version of the same red flag, quieter: the agency that will not tell you where it ranks. Ours is on the site — 151 reads across 27 queries over 13 consecutive days, 0 appearances — and the whole log is here.

04 — Timing

How long before any of this shows up

Longer than a paid campaign and on a clock you do not set. There are really three clocks running at once — how fast you publish, how fast the engines re-read the web, and how fast third parties pick you up — and only the first one is yours.

We measured our own domain daily to find out, and published the interval with the method and the flat stretches included rather than a tidy estimate. The measurement is here — it is the single most useful page to read before you agree to a contract length, and we are not going to paraphrase its numbers here, because a figure repeated without its method is how estimates turn into promises.

The practical consequence for a buyer: be suspicious of a three-month plan that is sold on citations and has no measurement in month one. Whatever the interval turns out to be for your category, you want the readings to have started before the work did.

05 — What August does

Concretely, in the order we do it

First the baseline: your queries, written the way your customers write them, read on a fixed method and from your city, before anything is changed. Then the citable layer on your own site — the plain answers to the questions people actually ask, with the structured data that lets a machine quote them. Then the part that usually moves the needle and usually gets skipped: the surfaces outside your site that answer engines assemble replies from. And a monthly reading that includes every zero.

We run exactly that on our own domain and report it in public, which is the only reason we are comfortable asking you to apply the five questions to us. Between September 25 and October 7, 2026 the panel recorded 151 reads across 27 queries and beaugust.co appeared in none of them. Within that total, two surfaces are counted separately and never compared to each other: 71 reads of Google's AI Overview block and 80 runs in ChatGPT. 0 appearances on either.

That is a bad number to publish and an honest one, and it is the same number we would show you about your own domain. The service page explains the scope and what gets delivered.

06 — Questions

The five questions again, to take with you

Same text as above, collapsed. Open one, read it in the call.

07 — Keep reading

The measurements behind this page

Every figure on this site comes with its date and its method. These are the pieces this one leans on.