How do you choose an SEO agency for AEO visibility?

To choose an SEO agency for AEO visibility, put four tests to it before you talk about scope or price, because each one is something a genuine practice can demonstrate in a first call and a repackaged one cannot. Most buyers arrive at this question already holding an SEO contract, and the honest answer is that some SEO agencies can do this work well and most cannot yet. The tests below tell you which kind you have.

Ask them to name the schema types they would add to your pages. An agency doing this work can say FAQPage, Article, Organization and BreadcrumbList without pausing, because structured data is how a machine reads what a page is claiming. Ask them to show you an AI answer, from any engine, that names one of their clients, with the date it was captured. Ask which engine each of their clients is weakest on, which is a question that only has an answer if somebody measured the engines separately. And ask what they would change on your pages in the first month, because an answer built entirely from backlinks and off-site work is the old playbook wearing a new label.

The reason those tests bite is that AI citation and Google ranking have come apart. ConvertMate's GEO Benchmark 2026 found that 83% of AI Overview citations point to pages outside the organic top 10, and that 68.7% of cited pages use a logical heading hierarchy. An agency that has only ever optimised for position is optimising for the wrong half of that finding, and it will keep reporting rankings that no longer predict whether an engine quotes you.

One warning about a promise that sounds reassuring and is not: an SEO agency that offers to guarantee AI citations is telling you it does not understand the mechanism. The same goes for a single blended "AI visibility score" with no per-engine breakdown underneath it. Both are covered in more detail in the five questions further down this page.

How do you test whether your current SEO agency can do AEO?

Test your current SEO agency with a thirty-day brief rather than a conversation, and ask for artefacts instead of opinions. A practice already doing this work produces four things as a matter of routine: a per-engine baseline, a page-level diagnosis, a list of the pages it would change first, and a sample report. A repackaged shop has to invent all four on the spot, which is visible in what comes back. Most buyers arrive at this question holding an SEO contract they would rather keep, so the brief is worth running before the shortlist, not after.

A thirty-day brief for an incumbent SEO agency, and how to read what comes back

What to ask for	What a real answer contains	What a repackaged answer contains
A baseline, engine by engine	Twenty to thirty of your buyers' prompts, run on ChatGPT, Google AI Mode, Perplexity and Claude separately and on more than one day, with the prompt wording and run dates in a file you keep	A screenshot of one ChatGPT answer, or a single blended visibility score with no prompt list behind it
A diagnosis at page level	Which of your URLs the AI agents fetched and passed over, and which were never fetched at all, read from your own server logs	A site-wide audit score and recommendations that would read identically for any client in any sector
The first month of page changes	Named URLs with the sections they would add or rewrite, in the words your buyers use rather than your product vocabulary	A calendar of new posts, or a link-building proposal with a monthly placement target
The reporting format	A sample report carrying per-engine numbers, the prompts where ground was lost, and the raw export the numbers came from	Dashboard access with one number on the front page and no way to export the runs
What they will not promise	A plain statement that citations are influenced rather than guaranteed, volunteered without being asked	A guaranteed placement, a citation count per month, or a promise about one named engine

One line on a deliverables list sorts the answers faster than the rest of the brief. If the first item offered is an llms.txt file, the agency is selling a signal almost nothing reads. Ahrefs examined the 137,210 domains in its web analytics product that received traffic in May 2026 and found that 97% of the sites publishing an llms.txt got no request for the file at all during that month, and Google's John Mueller wrote in June 2025 that no AI system currently uses it (Ahrefs, We Analyzed 137K Sites, published 15 June 2026, and Mueller on Bluesky, 17 June 2025; both figures confirmed from two independent accounts rather than from the sources themselves, 14 October 2026). The file is cheap and harmless to publish. It is not a programme, and what the four published llms.txt studies actually measured is the long version of why.

Should you keep an SEO agency that can do the page work but cannot measure AI visibility?

Keep it and buy the measurement separately, as long as the page work itself is good and somebody owns the brief that tells it what to change. This is the common case rather than the exception, and it is cheaper than a switch. An agency that already holds your style guide, your CMS access and your approvals process is expensive to replace, and what it is missing is a per-engine series rather than a skill.

Read the four answers together rather than scoring them one by one. An incumbent that returns a real baseline and a page-level diagnosis but has never produced a per-engine report is usually worth keeping, because the gap is a reporting habit. An incumbent whose answer arrives as new posts and backlink targets has told you what it thinks the work is, and that belief does not change because you ask twice. Whether to use an agency, a tool or your own team works through the three shapes, and what an AEO engagement includes is the list to read a returned brief against.

Should SEO and AEO sit with one agency or two?

You should keep SEO and AEO with one agency if you can find one that genuinely does both, and split them only when you cannot. The work overlaps heavily at the foundation. Crawlability, site structure, entity consistency and content quality serve Google ranking and AI citation at the same time, so two vendors working the same pages duplicate effort and eventually disagree about who owns the site.

Splitting has a real cost that buyers underestimate: nobody is accountable for the whole picture. When AI citations fall and rankings hold, or the reverse, two vendors each have a defensible story about why it was not them. You also end up running two measurement systems that do not reconcile, and paying for weekly calls to keep them from undoing each other's changes.

Splitting is defensible in two situations. The first is an SEO relationship that is working and that you do not want to disturb, where a specialist can take AI visibility as a separate workstream with a clear boundary: they touch content structure and extractability, the incumbent keeps technical SEO and paid search. The second is scale, where an in-house team already owns SEO and wants outside help only on the AI side. In both cases, write down which vendor owns page-level content changes before you start, because that is the exact seam where the fight happens.

Which kind of AEO provider fits which goal?

Match the provider to the goal before you start comparing firms, because the market sells AEO in five shapes and they are good at different jobs. Ask each firm what the acronym covers in their scope first, since answer engine optimization has no agreed definition and no single founding document, and two proposals using the word can describe different work. A buyer who wants a per-engine series and a programme of rewrites is shopping for something a PR firm cannot deliver, and a buyer whose pages have not been written yet is shopping for a writer rather than a measurement practice. The mismatch is a more common reason an engagement disappoints than the firm being weak.

Five provider shapes, and the goal each one actually serves

Provider shape	The goal it serves	Where the fit is narrow
Specialist AEO practice	You have the pages, and you need to know which engine drops you and why, then have the pages fixed	Throughput. These are small teams, and most carry no paid media or wider channel work
Full-service SEO agency	You want technical SEO, content and AI visibility on one contract, and the relationship already works	Measurement. Ask to see a per-engine series before assuming one exists
Content or editorial agency	The pages a buyer could be cited from do not exist yet, and somebody has to write them	Diagnosis. Volume without a baseline is writing in the dark, and nobody there reads server logs
Digital PR firm	You need third-party coverage and mentions on the pages engines already retrieve	Your own pages, which it does not touch, and which is where the fixable half usually sits
Consultant alongside your own team	You have writers and a developer, and what is missing is the brief, the prompt set and the reading of it	Delivery. Nothing ships in a month where your team has no hours

The label on the website does not tell you the shape, and one question usually does: what share of the retainer is measurement and writing hours, and who does them. A firm that cannot break that down has either not costed the work or is planning to subcontract it, and both are worth knowing before the scorecard further down this page. A tool plus your own team is the fifth option and the trade-offs sit in whether to use an agency, a tool or your own team, while what an AEO engagement includes is the list to read any proposal against.

Why AEO agencies can't be evaluated like SEO agencies

AEO agencies can't be evaluated like SEO agencies because the filters you would normally reach for do not exist yet. The usual way to vet a marketing agency — check their years in business, their case studies, their rankings — barely works for AEO, because the category is roughly two years old. Almost no one has a long track record, so "established" isn't an available filter. That creates a market where any SEO shop can rebrand as an "AEO agency" overnight, and a buyer has no easy way to tell a genuine practice from a repackaged one.

So the evaluation has to shift from credentials to method. The right questions aren't "how long have you done this" but "how exactly do you measure it, what can you show me, and where do the citations actually come from." The criteria below are built around the things a real practice does and a repackaged one can't fake.

How do you build the shortlist before you evaluate anyone?

Build the shortlist from four sources rather than from one search, because a single search for an AEO agency returns whoever wrote the best ranking page, and that is a different skill from the one you are buying. The tests further down this page sort a shortlist. They do not produce one, and most buyers arrive having let a search result produce theirs.

Start by asking the engines, since that is the behaviour you are paying someone to influence. Put the question a buyer like you would put, in your own words, to ChatGPT, Google AI Mode, Perplexity and Claude separately, and write down who each one names. Then ask again tomorrow. Two things fall out of that exercise. The firms named by more than one engine on more than one day are the ones with a real presence rather than a lucky retrieval, and the exercise shows you what the work looks like from the buying side, which makes the first sales call much easier to run. One query is not enough to be sure of anything, and how many prompts it takes to track AI visibility sets out how far that goes wrong when the sample is small.

Second, read the comparison pages, and read them as what they are. For a buying question the engines overwhelmingly retrieve ranking pages rather than agency sites, so those pages are shaping the answer whether or not they are any good, and most were written either by a firm on the list or by a publisher who sells the placement. That does not make them useless. The criteria a page states are usually honest even when the ordering is not, so take the criteria and discard the ranking. Ours is our comparison of AEO agencies for 2026, including where we fit and where we do not, and the mechanics of why these pages carry so much weight are in how to get into the listicles AI engines cite.

Third, look for the firms publishing method rather than opinion: a stated measurement approach, a sample of what a report contains, studies with their query sets and run counts written down. A firm that has published how it measures has committed to something you can hold it to, and it is also the cheapest signal to check, because the page either exists or it does not. Fourth, use the answers you already have. If an engine names a competitor of yours for your category, find out who does their AI work, and if nobody does, that tells you something more useful still.

One part of this is no longer guesswork, which changes what you should ask to see in a first meeting. Google reports AI-feature impressions for your own site in Search Console: traffic from AI features appears in the Performance report under the Web search type, and there is a separate generative AI report showing how often URLs from your site appeared in generative AI features across Search and Discover, which was rolling out to a subset of properties when we checked (Google Search Central, AI features and your website, and Search Central blog, generative AI performance reports, read 9 October 2026). It covers Google's surfaces only, so it tells you nothing about ChatGPT, Claude or Perplexity. Within that limit it is your data, in your own console, and an agency that cannot talk fluently about what it says is not measuring the channel it is selling.

How do you compare AEO agencies when every published ranking is written by one of them?

Compare AEO agencies on evidence you collect yourself, and read every published ranking as a statement of criteria rather than as a verdict, because in a category this young the firm that wrote the list is usually on it. The sources a buyer reaches for are not worthless. Each one is good for a narrow thing and unreliable for the thing buyers actually use it for, which is the order the names appear in.

What each source you would compare agencies with can and cannot settle

Source	What it can settle	What it cannot
A ranking published by an agency	The criteria the category judges on, and which firms treat each other as peers	The order, because the publisher is a candidate. Ours is no exception to that
A publisher or directory list	Which firms spend to be found, and occasionally the price of the position	Capability. Check whether the page marks paid placements at all before you read the ranking
A review platform profile	What named clients bought, in what sector, at what project size	Whether any of it was AI visibility work. Read the dates and what the reviewer says was delivered
The engines' own answers	Which firms hold a presence across several engines and more than one day	Capability in your category, and an engine will repeat a vendor's ranking of itself
A former client, or a competitor's supplier	What actually happened, with dates, including what the monthly report contained	Anything general. One engagement is one engagement

One piece of published evidence is worth carrying into the pitches, because it tells you which half of the citation surface a plan is aimed at. Ahrefs classified the top 1,000 pages ChatGPT cited in September 2025 from its own index and found that 32.3% were pages a marketer could realistically pitch or influence, the remainder being Wikipedia entries, company homepages and app store listings. Among the cited pages that had any search presence at all, 65.3% sat at a domain rating of 81 or above with a median of 90, while 67.3% had a URL rating of 10 or below with a median of 6 (Ahrefs, 67% of ChatGPT's Top 1,000 Citations Are Off-Limits to Marketers, published 28 October 2025, read at source 23 October 2026). It is one engine, one month and one vendor's index, the content types were sorted by a model rather than by hand, and the study is a year old, so it describes a shape rather than a constant.

The shape is still enough to put two proposals on different footings. Site-level authority separated those cited pages and page-level links barely did, so a plan built on links pointed at the page you want cited is buying the measure that distinguished them least, and a plan built on outreach alone is working inside roughly a third of the surface. Neither is disqualifying on its own. Ask which part of the plan is aimed at pages you own, which part is aimed at third parties, and what evidence the agency has that either one moved for a client it can name.

The criteria that actually matter

Six criteria do most of the sorting. They are arranged below as a table because the warning sign is usually easier to recognise than the good answer, and because a buyer sitting in a pitch needs something they can check against in real time rather than a philosophy. Each row is a claim you can test in a single question.

Criterion	What a real AEO practice does	Warning sign
Measurement	Measures all four engines (ChatGPT, Google AI Overview, Perplexity, Claude) separately, query by query, across repeated runs	Reports one blended "AI visibility score," or only checks ChatGPT
Evidence	Shows you your actual data — the queries where a competitor is named and you aren't — before pitching	Leads with promises and generic decks, no firm-specific data
Where the work happens	Fixes your own pages, where 73–89% of AI citations originate	Talks mostly about backlinks and off-site tactics borrowed from SEO
Honesty	States plainly that citations are influenced, not guaranteed	Guarantees citations or specific placements
Per-engine literacy	Knows ChatGPT leans on Bing, Claude on Brave, Perplexity on its own index	Treats "AI" as one undifferentiated thing
Conflict policy	Works with one client per competitive set, so they're not optimizing your rivals against you	Will take every firm in your market

How do you score two agencies against each other?

You score two agencies against each other by marking both on the same six criteria, out of two points each, weighting the marks, and treating measurement as a gate rather than as a weight. A shortlist conversation produces impressions rather than comparisons, and impressions favour whichever firm presented last. Writing the marks down while each call is still fresh is what makes the two proposals comparable at all, and it has the side effect of forcing you to notice which answers you never actually got.

A scorecard for two or three shortlisted agencies. Mark each criterion 0, 1 or 2.

Criterion and weight	Scores 2	Scores 1	Scores 0
Per-engine measurement (25, and a gate)	Reports every engine separately, prompt by prompt, across repeated runs, and showed you the format	Measures separately but reports a roll-up, with the detail available on request	One blended visibility score, or only ChatGPT
Firm-specific evidence before signing (20)	Put your own queries, your competitors and the capture dates in front of you unprompted	Showed another client's data, named and dated	Decks, category statistics and no finding about you
Where the work happens (20)	Named the page changes they would make in the first month, specifically enough to argue with	Described page work in general terms, with off-site work as the larger half	A plan built mostly on backlinks and placements
Honesty about limits (15)	Said plainly that citations are influenced rather than guaranteed, and described a bad month	Avoided guarantees but had no answer on what a bad month looks like	Guaranteed citations, placements or a timeline
Who writes the pages (10)	Their writers, with a named reviewer on your side and a sample you have read	Shared drafting, with your team expected to supply the substance	They deliver recommendations and the writing is yours
Conflict policy in writing (10)	One client per competitive set, with the set named from their own measurement	A policy by industry or by city, which is looser than it sounds	Will take any firm in your market

Multiply each mark by half its weight, so a 2 earns the full weight, a 1 earns half and a 0 earns nothing. That gives a weighted total out of 100 for each firm, and the worked example below shows why that total is the second thing you read rather than the first.

What does the scorecard look like applied to two proposals?

Applied to two proposals, the scorecard can separate them on something other than the total, which is the reason to write it out rather than hold it in your head. The two columns below are a worked example rather than real firms: a specialist practice that is thin on proof, against a well-run generalist that measures badly.

A worked example, not real firms. Marks are out of 2, and the points are the mark multiplied by half the weight.

Criterion (weight)	Agency A mark and points	Agency B mark and points
Per-engine measurement (25)	1, so 12.5	0, so 0
Firm-specific evidence (20)	1, so 10	2, so 20
Where the work happens (20)	2, so 20	2, so 20
Honesty about limits (15)	2, so 15	2, so 15
Who writes the pages (10)	1, so 5	2, so 10
Conflict policy (10)	1, so 5	2, so 10
Weighted total out of 100	67.5	75

Agency B wins the total by seven and a half points and should still lose the decision, because the zero it took is on the gate. A firm that reports one blended number cannot tell you which engine you are failing, which means it cannot tell you which fix to buy, and every strength further down the column is being spent without aim. That is the whole use of separating a gate from a weight: it stops a strong pitch on five criteria from covering a disqualifying answer on the sixth.

Two warnings about the arithmetic. A gap of five or eight points is inside the noise of who had the better call, so treat anything under about fifteen as a tie and go back for the answers you are missing rather than trusting the decimal. And a criterion you could not mark is not a zero, it is an unanswered question: write "not answered" in the cell, because three blank cells in one column is itself the finding, and it is the one most buyers talk themselves out of noticing.

Question 1 — Do they measure all four engines, separately?

Your visibility is not one number — it is four, one per engine, and they rarely agree. A firm can lead three engines and be invisible on the fourth, usually for a single fixable reason. An agency that reports a single blended score, or only checks ChatGPT, is hiding the engine-specific gaps that are the entire point of the work. Ask to see how they report — if it isn't broken out by engine and by query, they're measuring the wrong thing.

The single fixable reason is usually more mundane than buyers expect. ChatGPT's web results lean on the Bing index, so a company that is thin in Bing can rank normally on Google, Perplexity and Claude and still be absent from ChatGPT across an entire category. Our study of pharmaceutical process skid fabricators found exactly that pattern: the two most-cited firms in the clean-in-place category led three engines across nearly every query and disappeared on ChatGPT across almost all of them. A blended score would have shown those firms as healthy. The per-engine view showed a whole lane of queries recoverable by fixing one thing.

Question 2 — Can they show you data, or just promises?

An agency that can show you data rather than promises is the fastest filter available to you, and the one most firms fail. A real practice can usually show you your own situation before you sign — the specific queries where AI names a competitor instead of you — because they've run your market. Data depth is what separates a practice that has actually measured the landscape from one selling a generic service. If a prospective agency can only talk in generalities and can't put a single firm-specific finding in front of you, they haven't done the measurement that the work depends on.

What you are looking for is specific enough to describe. A named competitor, a named engine, a buyer query in the words your customers actually use, and the date it was captured. Anything softer than that is a slide, not a finding. Be wary too of screenshots with no timestamp: engine answers move week to week, and an undated capture could be from a run six months ago that no longer reproduces. A practice confident in its method will happily re-run a query in front of you. Apply the same scepticism to the answer itself, because an engine will repeat a vendor's own ranking of itself without marking where the claim came from.

Question 3 — Do they fix your pages, or sell you backlinks?

Across measured categories, 73–89% of AI citations point to companies' own pages rather than directories or third-party sites. That means the primary lever is your own content — its structure, specificity, and language. An agency whose plan is mostly backlink-building and off-site SEO tactics is applying the old playbook to a problem it doesn't fit. The work that moves AI citation happens largely on your own pages; make sure that's where their effort goes. (For the underlying mechanics, see how AEO differs from SEO.)

Page-level work is concrete, so a capable agency can describe it before it has touched your site. Rewriting a heading into the question a buyer actually asks. Putting a direct answer in the first sentence under that heading rather than three paragraphs down. Adding a comparison table where you currently have prose. Making sure the claim an engine would want to quote sits in a sentence that still makes sense when it is lifted out of the page. Stating specifications, standards and configurations explicitly instead of describing them as bespoke solutions. If a prospective agency cannot name work of that shape for your pages, it is planning to sell you something else.

Question 4 — Are they honest about what can't be guaranteed?

An honest agency will tell you plainly that nothing here can be guaranteed, because AI citation is influenced rather than controlled. The engines make the final call and they change their behaviour without notice. Any agency that guarantees citations or specific placements is either misunderstanding the mechanism or overselling it. A trustworthy practice is explicit about this: they remove the structural reasons you're skipped and improve your odds, and they report honestly including where they haven't moved your position yet. Guarantees are a warning sign, not a reassurance.

There is a version of this that is easy to miss, because it sounds like confidence rather than a promise. An agency that shows you only the queries where you improved, and quietly drops the ones where a competitor took your place, is making the same mistake with better manners. Ask what their reporting looks like in a bad month. A practice that has an answer to that question has run enough campaigns to have had one, and a practice that does not is either new or editing the picture. Neither is disqualifying on its own, but you want to know which you are buying.

Question 5 — Will they work with your competitor too?

Ask whether they will work with your competitor too, and treat a vague answer as a no. Because AI answers name a short list, helping two competing firms in the same market is a conflict — the agency would be optimizing rivals against each other. A serious practice works with one client per competitive set per market. If a prospective agency is willing to take every firm in your city, ask what happens to your position when they sign the firm down the street.

Get the answer into the contract, and define the boundary yourself rather than accepting theirs. "One client per industry" is close to meaningless for a firm that competes in three categories in two cities. What you want named is the competitive set: the companies that appear alongside you when an engine answers the queries you care about, which is a list the agency should already be able to produce from its own measurement. Agree what happens if one of them approaches the agency later, and agree it before you sign, because the conversation is much harder once the retainer is running.

How do you check an agency's own AI visibility before you hire them?

You check an agency's own AI visibility the way you would check a client's, by putting the category question to each engine in a buyer's words and reading which sources the answer names. Ask ChatGPT, Claude, Perplexity and Google AI Mode something like "who should I hire to get my company cited in AI answers", then ask the narrower version your own buyer would use. Write down whether the agency appears, and more usefully, which page the engine cited when it did.

The page matters more than the mention. An agency named from a third party's listicle has been picked up by somebody else's page, which tells you about that listicle. An agency named with its own page cited underneath has done on its own domain the thing you are about to pay it to do on yours. That second result is the one worth something, and it is the one a repackaged SEO practice almost never has.

What to test	A good result	What it does not prove
The category question, asked of all four engines	The firm is named, and at least one engine cites a page on the firm's own domain	That they can do it in your category, which may be harder or easier than theirs
A client page they say they worked on	The client is cited for the query the page targets, and the page reads as an answer rather than a brochure	That they caused it. Ask when the page changed and what it looked like before
The same query, run again three days later	The firm or the client page holds across runs rather than appearing once	Nothing, but a single run proves nothing either, which is the point of repeating it
Their own pages against the mechanics they sell	Server-rendered HTML, question-shaped headings, sourced figures, a visible date	That the content is any good. It only shows they practise the technical half

Run each query more than once before you conclude anything, because a single answer is a sample rather than a finding. GetMentions put the same 2,398 questions to ChatGPT, Google AI Mode, Perplexity and Gemini every day for a week, logged 530,875 citations, and found that 69% of the sources behind a typical answer changed from one day to the next, with 84% of the sources cited for a question used by only one of the four engines (GetMentions, "AI Citation Volatility: A 530,875-Citation Study", checked 28 September 2026). At that rate of turnover, an agency can be absent from Tuesday's answer and present in Friday's without anything having changed about the agency.

Treat a weak result as a question rather than a verdict. A three-person practice doing good work for industrial manufacturers can be invisible on "best AEO agency" while owning the queries its clients care about, because the category term is contested by firms whose whole business is ranking for it. What should worry you is the reverse: a firm that is highly visible on the category term and cannot show you a client page that is cited for anything.

What should you ask an agency's references?

Ask an agency's references for the things the agency cannot stage: when the citations appeared, what the reporting looked like in a bad month, and who did the writing. A reference call that stays on whether the team was pleasant to work with has told you nothing you could not have guessed.

Four questions do most of the work. First, how long was it before anything moved, and what moved first. A real answer distinguishes between the first citation and the first citation that mattered, and it usually lands in months rather than weeks. Second, what did the monthly report actually contain, and did it ever say that a number went down. An agency that has never reported a decline to a client is either very lucky or not measuring. Third, who wrote the page copy, and did your team have to rewrite it. This is where a repackaged practice shows, because the fix for a citation problem is usually a rewrite and a rewrite needs somebody who understands the product. Fourth, what happened at renewal, and what did you decide.

The reference worth calling is the one whose engagement ended. A current client on a live retainer has a reason to be generous, and the agency chose them for it. A former client has no stake either way, and if the agency cannot offer one at all after two years of trading, that absence is the answer. When you ask for a dated citation as proof, our note on how to tell whether AEO is working sets out what a defensible before-and-after looks like, so you can check the reference's account against it.

What should an AEO engagement cost, and what changes the price?

An AEO engagement costs between $3,000 and $15,000 a month across most of the market in 2026. Mid-market programs cluster at $3,000 to $8,000 a month, and enterprise scopes carrying heavy off-site work start near $15,000, per the 2026 AEO and GEO pricing guide from Digital Elevator, checked 16 September 2026. Four things move a quote within that range, and all four are worth asking about before you compare two proposals.

Scope	Typical monthly range, 2026	What that usually buys
Mid-market program	$3,000 to $8,000	A defined page set, per-engine measurement, and the on-page work to fix it
Enterprise program	From about $15,000	Large page counts, heavy off-site work, frequent prompt auditing

Market ranges, not our rates. Both figures are from the Digital Elevator 2026 pricing guide linked above, checked 16 September 2026. Pricing in a two-year-old category moves quickly, so treat these as accurate on that date.

The first is the number of pages in scope, because the work is per page and a hundred-page site is not five times a twenty-page site. Then whether off-site work is included, which is the expensive half and the half nobody can guarantee. How often the prompt set is re-run, because measurement is a recurring cost rather than a one-off. And whether the retainer includes the writing, which is where most of the labour actually sits: an agency that measures and hands you a list is selling a smaller thing than one that measures and then rewrites the pages.

Treat a very low number as information rather than a bargain. Our own read of this market, and it is a judgement rather than a published figure, is that below roughly $1,500 a month there is not enough labour in the scope to do the work described on this page, so what is being sold is usually a tracker subscription with a report attached. That is a real product and it is worth its price, but it is not an AEO program, and the comparison of AEO tools is the right page to read before paying agency rates for it.

Price is also a weak quality signal here, weaker than it would be in an established category. AEO is roughly two years old, so a high retainer buys you no evidence of competence on its own, and the five questions above remain the better filter. What the number is genuinely useful for is checking that the scope is honest: an agency quoting mid-market money for enterprise-sized promises has either not costed the work or is not planning to do it.

You will notice we do not publish our own figure. Scope, page count and how much of the writing we take on change it enough that a headline number would mislead more than it helped, so we quote after the assessment rather than before it.

What should the monthly report contain?

A monthly report should contain six things, and five of them are absent from most of the reports we see. The prompts and the engines, each result dated, so the series is comparable with last month's. The prompts where you lost ground, named rather than netted off against the wins. Every URL the engines cited behind those answers, because the page that was quoted is the finding and whether you appeared is only the headline. What changed on your site during the period, and on which date, since nothing in the measurement can be attributed without it. Where the competitive set stands on the same prompts. And what the agency is doing next, with what it is waiting on from you.

The fourth of those is the one that separates a report from a dashboard screenshot. Citation movement has a lag, and a page edited on the 8th cannot be judged by a run on the 10th, so a report that lists measurement without listing changes leaves both sides free to claim the movement or disown it. A change log with dates, read against a prompt series with dates, is the only arrangement in which an argument about cause can be settled at all. Ask for the two in one document rather than in two systems.

Two things are worth refusing outright. A single blended score with no per-engine breakdown, for the reason the first question on this page sets out: it cannot locate a failure, so it cannot direct any work. And an answer screenshot with no capture date, because engine answers move week to week and an undated capture cannot be checked or re-run. Our note on how to tell whether your AEO is working sets out what a defensible before-and-after looks like, and it is a reasonable standard to hold a monthly report to.

What should be in the contract before you sign?

The contract should name six things: the prompt set, the raw measurement history and the format you receive it in, who may edit your pages, the competitive set, what a report looks like in a month where you lost ground, and what you keep if the relationship ends. All six are cheap to agree before a retainer starts and expensive to argue about afterwards, because the measurement is only worth something as an unbroken series and the pages are yours either way.

Start with the prompt set, since it is the asset the rest of the work is built on. A citation series stays comparable only while the wording stays fixed, so an agency that rewrites prompts between reports has broken the comparison that shows whether anything improved, and one that will not hand the list over has made your history unportable. Ask for the prompts in a file you keep, and agree that the list grows by addition rather than by editing what is already in it. The design behind a usable prompt set is covered in our guide to tracking AI visibility across the four engines.

Then ask for the runs themselves rather than for access to the dashboard that displays them. A decline is only visible as a difference between runs, so a program that changes hands with no exports restarts from no baseline, and the first quarter goes on rebuilding a picture you have already paid for once. A line specifying per-run exports, with the prompt, the engine, the date and every cited URL, costs an agency nothing if it is measuring properly. Our guide to measuring AI citation decay between runs sets out what those exports need to hold to still be readable a year later, and the SIGNALS methodology sets the cadence we run them on.

Term	What to write down	What it prevents
Prompt set	The exact prompts, engine by engine, in a file you receive at the reporting cadence	A rewritten prompt that resets the series and hides a decline
Measurement history	Per-run exports carrying the prompt, the engine, the date and every cited URL	Losing the baseline when the engagement or the agency changes
Page access	Who edits, who approves, and whether changes go through staging	Two vendors overwriting each other, or an unreviewed claim going live
Competitive set	The named companies that appear beside you in the answers, taken from the agency's own measurement	A conflict arriving later under a vague "one client per industry" clause
Reporting in a bad month	That reports carry the prompts where you lost ground, not only the wins	A report edited into a success story
Exit	What you keep: the prompts, the history, the page changes and any schema added to your site	Paying for work that leaves with the agency

An agency that resists these terms has told you something useful before you have spent anything. None of them asks for a guaranteed outcome, and none is unusual in an older category, where a contract that did not say who owns the tracking or who may edit the site would look careless. AEO is young enough that the equivalent clauses are often missing, which is a reason to write them yourself rather than evidence that they do not matter. Whether this work belongs with an agency at all is a separate question, and our comparison of an agency against a tool and an in-house team is the page for that one.

Who should be allowed to edit your pages?

Page access is the seam where the friction usually happens, because this work is page editing and somebody else may already hold the keys. Write down who edits, who approves, and whether changes go through staging, particularly if an incumbent SEO agency is working the same templates. The clause protects you in the other direction too: an agency that can publish to your site unreviewed can put a claim on a commercial page that nobody in your company has checked.

The arrangement that works in most companies is that the agency drafts, somebody on your side approves, and the change ships through the same review path as any other content. Ask what happens when a fix is urgent, because that is the moment the path gets skipped, and ask who holds the credentials once the engagement ends. An agency that wants direct publishing rights from the first week is not disqualified by that, but it should come with a named approver on your side and a log of what changed and when. A firm doing this work at any scale already keeps that log, because it needs to know which change preceded which movement in the measurement.

What else do people ask about choosing an AEO agency?

What else do people ask about cost, contracts and reporting?

Related guides

• AEO agency vs. tool vs. in-house — which delivery model fits, compared side by side.
• Best AEO tools in 2026 — the tool side of the comparison set, ranked.
• AEO vs. SEO — why the old agency-vetting playbook doesn't transfer.
• How to track AI visibility — measure where you stand before you hire anyone.
• The 7 SIGNALS dimensions and the research methodology behind them.