Your agency guarantees AI citations. What exactly are you buying?

TL;DR: A “guarantee” of AI citations can promise more than its evidence supports. Publishing volume, selected screenshots and successful manipulation tests do not establish dependable recommendations for your buyers. Before signing, check what would count as failure, how results are tested and whether the cited page supports the buyer's next step.

The pitch

I received a cold email earlier this month that described my business accurately. It said Xyra Foundry can show clients where they are missing from AI answers, but suggested there was no next step once the gap was found. The sender offered the execution layer: getting brands recommended and cited by ChatGPT and Claude through more than 1,500 sites it described as high-authority sources that AI trusts.

I read it as a paid-placement offer presented in the language of AI visibility. The email did not establish how those sites were selected, how placements were disclosed or what counted as being recommended. That was what I wanted to understand before taking the promise at face value.

Three ways an AI citation “guarantee” can mislead you

An agency can influence some inputs to an AI answer, but it does not control the final response. A “guarantee” may promise dependable recommendations while its evidence establishes something much narrower.

A guarantee can be met without any of these showing dependable recommendations.

1. Volume sold as influence

An agency promises hundreds of posts, pages or comments, presenting the volume as a route to guaranteed citations. Counting what it publishes tells you nothing about whether AI recommends your business for relevant buyer questions. The output is measurable; the promised influence still needs evidence.

2. Selected wins sold as repeatability

A screenshot shows your brand appearing in an answer. It does not show how often the brand was absent across other prompts or runs. If the agency chooses its examples after the work is done, it can present selected successes as dependable visibility. One answer proves one appearance, not a result your buyers can reliably expect.

3. Manipulation sold as dependable visibility

Hidden instructions or crafted text can influence AI outputs under tested conditions, as the studies in Part 1 show. That does not establish reliable recommendations across engines, buyer questions or future versions. A successful manipulation test is not evidence that an agency controls your visibility.

Before paying for a “guarantee”, ask what would count as failure. If the agency can fulfil it by publishing enough material or finding one favourable answer, the promise may be met while the commercial result you expected remains unproven.

What a citation actually tells you

A citation is a source link displayed with an AI answer. It identifies a source presented for that response; it does not, by itself, establish an endorsement or accurate support for the claim.

It does not distinguish a mention from a recommendation

A brand mention, an explicit recommendation and a source link are different measures of AI visibility and impact. A dashboard may count appearances equally even when one is a passing mention and another names the business as the best fit. Ask which observation the agency is reporting.

It does not establish buyer demand

A citation in an answer to a question your customers never ask tells you little about commercial opportunity. A prompt set should reflect recognisable buyer questions. Even then, observed answers are not a measurement of how many real buyers saw them.

One answer per engine, one brief, collected once. It shows a possibility, not how often it happens.

It may send the buyer to somebody else's website

In my exploratory accommodation study, I collected one answer from each of four AI engines to the same brief. Three returned shortlists. Those contained nine recommendations covering seven properties. For one shared property, two engines linked different destinations: a curator's listing and the operator's own page.

An aggregate citation count alone hides that distinction. The property was recommended through two different websites, each with its own presentation of information and booking entry points.

This was a one-off sample with incomplete run metadata. It does not measure frequency, citation lifespan or conversion. A supplementary screenshot review established visible page differences, not which page performed better.

It gives you no fixed lifespan

A source can remain useful for years, change, move or be removed. In a published account, a ReachLLM marketer described posts cited within a day that Reddit later removed. That is a reported case, not a general expiry date for citations.

Source removal and disappearance from AI answers may happen at different times. Ask what you control, what depends on continued payment and how the agency monitors changes. Owning a page gives you control over its content, not over whether an engine cites it.

Questions to ask an agency before you buy AI visibility

Ask the agency

1. Define the outcome. Is the target a brand mention, an explicit recommendation, a linked source, referral traffic or an enquiry? Which part is guaranteed, and what happens if it is not delivered?

   
   

2. Agree the measurement. Which engines, interfaces or APIs, prompts, modes and dates will be used? Fix the prompt set before work starts. Ask for raw answers, cited URLs and a baseline, rather than selected screenshots or aggregate volume alone.

   
   

3. Identify the sources. Which websites will carry the content, who controls them and which placements are paid? Ask for source URLs when published and the disclosure arrangements. Commissioning deceptive reviews can expose both client and agency to liability.

   
   

4. Define the continuing commitment. What happens if a page is changed or removed, a target is missed or payment stops? Ask which content remains live, what maintenance ends and what evidence supports any claim about durability.

Check on your side

1. Check what buyers are being told. If an answer merges two services, repeats an old price or misstates a policy, investigate the source and correct the information you control before paying to amplify it.

   
   

2. Check who the prompts represent. Do they describe the buyers, needs and decisions your business serves? Being visible for an easy prompt is of limited use if the question is unrelated to what you sell.

   
   

3. Check whether your website finishes the job. An AI answer may have shaped a visitor's expectations before they arrive. Can the landing page substantiate those expectations, explain important conditions and provide an appropriate next step?

   
   

If that information is difficult to find, the visit may stall. Traffic counts alone will not explain why.

Before signing, establish what the agency will deliver, where it will appear and how success will be measured. Then check the page a buyer would reach: can they verify the recommendation and take the next step? That is where visibility becomes a commercial opportunity.

My AI readiness and buyer journey audit examines that connection: how AI represents your business, which page the buyer reaches and whether it supports their next step.

---

Frequently asked questions

Can an agency guarantee AI citations?

An agency can influence some inputs to an AI answer, but it does not control the final response. A guarantee can therefore promise more than its evidence supports. Publishing volume, selected screenshots and successful manipulation tests do not establish dependable recommendations for your buyers.

What is an AI citation?

An AI citation is a source link displayed with an AI answer. It identifies a source presented for that response. By itself it does not establish an endorsement or accurate support for the claim.

What should an AI citation guarantee specify?

It should define the outcome, such as a brand mention, an explicit recommendation, a linked source, referral traffic or an enquiry, and state which part is guaranteed and what happens if it is not delivered. It should also fix the engines, prompts, modes and dates before work starts, with raw answers and cited URLs as the evidence.

Does a screenshot of an AI answer prove an agency's work?

A screenshot shows one appearance in one answer. It does not show how often the brand was absent across other prompts or runs, and an agency that chooses its examples after the work is done can present selected successes as dependable visibility.

How long does an AI citation last?

There is no fixed lifespan. A source can remain useful for years, change, move or be removed, and its removal and its disappearance from AI answers may happen at different times. Ask what you control, what depends on continued payment and how the agency monitors changes.

Can an AI citation send buyers to a different website from yours?

Yes. In one exploratory accommodation sample, two engines recommended the same property through different websites, a curator's listing and the operator's own page. The sample was one answer from each of four engines, so it shows a possibility and says nothing about how often it happens.

If you are weighing up a guarantee, book a discovery call and we can go through what it would commit the agency to deliver, and how you would know.