How to choose an AI visibility agency, and the one question that sorts them

Four kinds of business are selling AI visibility this year, and their pitches are almost impossible to tell apart. Only one kind can put a word on the third-party pages an AI answer is assembled from. The question that sorts the category takes eight seconds to ask: can you place the content, or can you only report on it?

This is the shortlist conversation, for the planner comparing three suppliers whose decks carry the same six words. Avid sits in one of the four groups below, so read this knowing we have a side in it. The questions work on us as well as on anyone else.

Four kinds of supplier, and only one can publish

The software vendor. A subscription that runs a set of buyer questions across the models every day and reports where you are named and which sources the models used. Buy it for the read, which is worth having. It will tell you the answer about your brand is wrong, and unless something else changes it will still be telling you that in six months.

The search agency. Markup, question-shaped headings, clean pages, topic clusters, a machine-readable file at your root. Nearly all of it is work on your own website, most of it is sound hygiene, and several of the tactics doing the rounds this year are contested rather than settled. Keep the search agency. Ask it which of its deliverables sit on your own domain, and how much of the answer it believes those pages decide.

The public relations agency. Closer to the mechanism, because it works on other people's pages and knows why a third party saying something counts for more than you saying it. Its currency is the news agenda, which is also its limit. You cannot brief a news desk to carry one claim across four unaffiliated titles inside one eight-week window, and you should not want to.

The media partner. Commissions content with publishers, briefs the claim, and pays for it the way media is paid for. The article is written and edited inside the publisher's own editorial process, which is what makes it worth citing and what keeps it on the right side of platform policy on third-party content. This is branded content, a channel that existed long before AI answers did, now doing a job nobody built it for.

A supplier working only on your domain is working on a tenth of the answer

The pages that decide your answer are mostly not yours. Foundation Marketing and AirOps read 57.2 million citations across 5.1 million AI responses between December 2025 and February 2026, and found that 90% of the sources AI cites when answering are ones the brand does not control; only 10.15% of citations go to brand-owned domains (Foundation Marketing and AirOps, May 2026, market not stated).

Treat that as a budget instruction rather than a headline. A supplier whose entire deliverable list sits on your own domain is working on the tenth of the answer you already own. The nine tenths that decide whether you are named at all belong to somebody else, and getting a brand into them is a commissioning job, not a reporting job. That is the whole reason the category is confusing: the measurement and the remedy are sold by different kinds of company, and only the measurement half is easy to demonstrate in a pitch.

Five questions for the first meeting

The useful part is not the question but what a real answer sounds like.

• Which sources decide the answers in our category, and can you name them? A good answer names publications, pages and the questions they show up against. A weak one describes a type of source: authoritative sites, trusted domains, high quality third parties. If a supplier cannot name the pages, it has not read your category, it has read the category in general.
• What will you publish, where will it sit, and who signs it off? You want titles, a number of pieces, a window, and a straight answer on who writes it. The right answer is that the publisher writes it to a brief and keeps editorial control, because a piece the publisher would not have run is worth less to a model and less to a reader.
• What will you read before we start, and will you run the same read at the end? A supplier that measures only after the work has run cannot show you movement, and one that changes the question set between reads has given itself a way to look good. Set the objective before anything is commissioned, which is the same discipline that keeps a content campaign out of trouble in every other channel.
• What are you doing that our own team cannot do on our own site? This is the question most suppliers have no answer to. If the plan is a content calendar, a knowledge base and better markup, your team can run that, and you are paying an agency margin for supervision.
• What happens when the answer about us is wrong? Not missing, wrong. A model repeating an out-of-date price or a rival's framing of your category is what brought most brands to this conversation, and the fix is to put a better sourced account of the same subject where the model is already looking.

Four answers that should end the meeting

• A promise that you will be named. Nobody controls the model, the answer changes week to week, and a supplier who will promise a position in an AI answer will promise anything else in the contract too. The honest commitment is movement on a stated set of questions, measured the same way at both ends.
• A plan built on raising your authority score. Surfer looked at around five million unique citation source URLs across 20,000 prompts over three months and found close to no correlation between a domain's authority score and whether AI cites it, with Spearman correlations running from -0.073 to +0.026 (Surfer, July 2026, market not stated). Being a big name is not the ticket in, which is good news if you are the challenger and worth knowing if you are the incumbent paying for the assumption.
• One blended visibility score. A buyer asks several different kinds of question on the way to a decision, and they need different measures. Blending a strong result on people searching your brand name with a weak result on people describing a problem produces a number that always looks healthier than the business is. Ask for the questions where you are still being chosen, read on their own.
• A twelve-month retainer whose first quarter is set-up. If nothing is published before month four, you have bought a research project. You should be able to see something in market inside the first two months and read the effect of it before you commit to the next phase.

The first commitment is one topic, one window, three kinds of publisher

It should be small enough for one person to sign off, and shaped so it proves something either way. In practice that means one topic rather than a spread of them, nine to thirteen articles across three deliberately different kinds of publisher, all of them unaffiliated and all live in the same window, and the same read run at the start and at the end so the movement is measured against where you began. One article on one title is not a fair trial of the idea, because a model weighing several independent accounts of a subject is doing something a single placement cannot show you. That shape is what an AI visibility campaign is built around, and any supplier should be able to describe theirs in the same amount of detail.

Ask, too, what you keep. Articles stay live and keep being read by models long after a media flight is over, which is the part of this that behaves like an asset rather than a burst. It is also why the perception work is slower and stickier than a promotion: a brand that shifted how it was described in third-party coverage did not do it with one piece in one week.

It sits with the team that already buys partnerships

It is bought as media, not as a search retainer. The deliverable is commissioned editorial on somebody else's title, priced and booked the way any other publisher activation is, which puts it with the team that already runs partnerships and gives it a natural home on the plan. Agencies reading this are in both seats at once: you will be asked to recommend a supplier and you may be the one delivering, and the same five questions sort the market either way. Our workspace runs exactly that comparison: one brief out, several publisher responses back, side by side.

Score every supplier on one thing: can they publish

Take the five questions into the meetings you already have booked, and score the answers on one thing: whether the supplier can name the pages that decide your category and then publish on them. If you want the first half of that answered before you spend anything, request an AI Visibility Report and you will see which questions your brand shows up against today, what gets said, and which sources the models are drawing on. Then ask every supplier on your list what they would do about it, and hire the one whose answer includes a publisher.