What an AI SEO agency does
An AI SEO agency works to get your business named, and your pages cited, in the answers that ChatGPT, Perplexity, Claude, Gemini and Google's AI answers write when your customers ask for help. The same service is sold as an AEO agency, a GEO agency, an LLM SEO agency, an AI SEO company or an AI search optimization agency. A real engagement contains four kinds of work: measuring which questions name you, making sure AI crawlers can read your site, publishing pages that answer buyer questions plainly, and building your presence on sites you do not own.
This page covers what each of those four should include, how to tell an agency that does them from one that renamed its old menu, the questions and warning signs that separate the two, and how we do this work at RunOctopus. If you want the definition of the discipline itself before judging suppliers, start with our page on what AI SEO is and what changes from regular SEO.
The name exists because the shape of the result changed. Ten blue links had room for ten businesses, and being eighth still put you on the page. A written answer names three or four and stops. There is no eleventh place and no partial credit for nearly making it. What did not change is the step underneath. An assistant asked something current runs a search, reads what comes back and writes from those pages. So most of what decides which pages rank still decides which pages an assistant ever sees.
That is why the labels matter so little. Answer engine optimization, generative engine optimization and LLM SEO describe roughly the same activity. Which one an agency puts on its website tells you about its marketing, not its method. The same four kinds of work sit underneath every one of them, and the rest of this page is about checking that they are really there.
The four parts of the work, and how fast each one moves
Strip away the naming and almost every genuine engagement is some mix of four activities. They differ enormously in cost, in speed and in how much anyone controls them. A proposal that does not separate them is describing a category, not a service.
Measurement
A fixed set of questions written in the words your buyers use, put to several assistants on a schedule, with the full answers saved. It is quick to set up and it changes nothing by itself. It is also the only thing that lets anyone show later that the rest of the work moved anything.
Technical access
Confirming that AI crawlers can read your pages: what your robots.txt file allows, whether your security layer challenges them, and whether your pages show their text without heavy scripts. This is binary, it blocks everything else when it fails, and it is usually a few hours of work. Presented as a months-long program, it is padding.
Content
Pages that answer the questions buyers ask, with the answer in the first paragraph and claims specific enough to quote. This is where most of the billable time goes and where quality varies most between suppliers.
Presence you do not own
Being described accurately on review sites, directories, forums and trade or local press. It is the slowest part, the least controllable, and often the one that decides whether you get retrieved at all. Our page on how LLMs choose sources explains why assistants lean on these outside voices more than on what a business says about itself.
Buyers usually misjudge the proportions, and usually in the same direction. Technical access feels like the serious, expert part, and it is a few hours. Content feels like the commodity part, and it is most of the value. Presence elsewhere feels like something you either have or do not, and it is the slowest thing to build. An engagement weighted heavily toward technical work has often run out of things to do.
The timelines differ just as much. Access fixes take days. Rewriting the openings of your important pages takes weeks, and most of that is approval time, not writing time. Building real depth on one subject takes quarters and does not compress. Being referenced by others has no timeline at all, because it is a consequence of the other work, not a task. A proposal that puts all four on one schedule will report the fast half and quietly defer the rest.
Ask how many hours go to each of the four parts. An agency that cannot answer has not scoped your work, whatever the proposal says.
What you should receive, in order
Whatever an agency calls its packages, a sensible engagement produces five things in a sequence where each one makes the next judgeable.
1. A baseline. Your buyer questions run across the assistants you care about, with the full answers saved, the businesses named and the date of each run. It is the most important deliverable and the one most often skipped, because the numbers usually look bad and nobody enjoys opening with that.
2. Crawler access. A written note saying which AI crawlers were checked, by name, on your site and on the network layer in front of it, and what was fixed. It belongs in week one.
3. Answer-first pages. Each page aimed at a real buyer question where a competitor is named and you are not, with the answer in a sentence an assistant can lift.
4. Work in outside sources. A named list of the review sites, directories, forums and publications that matter in your category, with what was checked and corrected in each.
5. A repeat measurement. The same questions, the same assistants, run again on a schedule, with a report that says what moved and what did not.
If an engagement skips the first step or the last, nothing in the middle can be judged.
What a good first ninety days looks like
The first two weeks should be almost entirely diagnostic: access read and reported, a baseline taken, and an honest statement of which of the four parts is actually holding you back. If month one produces a content calendar instead, the diagnosis was skipped and the plan is generic.
Weeks three to eight should be the fast half: access corrected if needed and the openings of your commercially important pages rewritten so they answer their own question at once. An early result, if there is one, shows up here in the assistants that look things up live.
Weeks nine to twelve should start the slow half, which means committing to one subject and beginning a connected set of pages that covers it properly. One subject, not four. By the end of ninety days you should hold a baseline, two later readings against it, a list of access fixes, a set of rewritten pages and a subject chosen with reasons.
Two kinds of agency use the same label
Almost every agency that sold search work last year sells AI search work this year. That is not a scandal. Much of the underlying craft carries over, and honest agencies say so. The problem is that you cannot tell the honest ones from the rest by reading their websites, because the websites all use the same words. Sit through enough sales calls and suppliers fall into two groups.
The renamed menu
This agency already sold search optimization, content and link building. It added a line for AI search, usually by relabeling existing services. The audit now mentions AI. The monthly report has a new page with a screenshot from ChatGPT. The work you receive is the work you would have received last year. That is not automatically bad, because fixing crawl problems and thin pages helps in both worlds. The issue is paying a premium for a label. If the deliverables match the old menu, the price should match the old menu too.
The agency built around answers
This supplier starts from the question, not the keyword. It can tell you which questions your buyers ask, which assistants answer them, what each assistant said last month and which businesses it named. Its work plan is built around the gaps in those answers. You can tell this kind by what it asks you: how your customers phrase their worries, what they compare you against and where they read before they buy. A renamed menu asks for your analytics login and your keyword list.
A useful test is to ask what the agency would do differently for AI answers than for ordinary search. The honest answer is a short list: write so a passage can be quoted, measure by asking assistants instead of watching traffic, and put more effort into being referenced elsewhere. A long list deserves a request for specifics. An empty one means the price should reflect it.
Ask for evidence, and know what counts
Every agency will show you something to prove it works. Not all proof is equal, and most buyers treat it as if it were. Rank it from weakest to strongest.
Logos and testimonials show the agency has had clients. They say nothing about whether any client was named more often afterward.
Rankings and traffic charts prove skill at ordinary search, which is fine, but a client can rank first and never be named in an answer, and the reverse also happens.
A single screenshot is one draw from a variable process. Ask the same question twice and the answer can change, so one screenshot may be typical or the best of twenty tries.
A repeated measurement is real proof of a trend: the same questions, run on a schedule, against the same assistants, with the full answers saved. Ask to see one real series with the client's name removed.
Before and after pairs on the same question are the strongest: the question, the answer before, the work done and the answer after, with dates. Ask for three, and ask for one where the work did not move anything. Every honest agency has one, and how it explains the miss tells you more than any success story.
How to read a case study without being fooled
Check the starting point, because a client who began near zero shows a huge percentage from a tiny absolute change. Check the time frame, because two weeks may be noise and three months is more believable. Check what else changed, because a product launch, a sale or a news story can be claimed as the agency's work. And check that the client is real and reachable. A reference who describes the working relationship is worth more than a polished page.
You can take a rough baseline yourself before talking to anyone. Our guide to the AI visibility checker explains a hand method that takes an afternoon, and our walkthrough on tracking brand mentions in ChatGPT shows how to keep it running. Having done it once, you will recognize quickly whether an agency is describing the real thing.
Agency, consultant, expert or yourself
A one-person consultancy, a mid-size agency and a done-for-you service are different tools. Matching the tool to the job saves money, and the deciding question is simple: who will write and publish the pages after the advice arrives?
When a consultant is enough
A consultant measures how assistants answer your buyer questions, finds the gaps and writes a plan ordered by cost and impact. Most do not build the changes. If you have someone who writes well, someone who can edit the site and a modest number of problems, that is good value and it keeps the knowledge in house. Ask the consultant to estimate the hours the plan needs from your side. A good one can say "about two days a week of writing for three months." The real cost of advice is the advice plus the work it creates.
When it is not
If nobody on your side can write and publish, a plan will sit in a folder. A consultant's report that nobody acts on is the most common way this money is wasted. Volume, follow-through, outside work and continuous measurement are the four things a consultant working alone usually cannot give you, and they are exactly what an agency or done-for-you service is for.
The four skills a real specialist has
The field is a couple of years old, so nobody has twenty years in it, and there is no certification worth relying on. Whoever you hire, test four skills with a real question from your business. Do they understand retrieval, meaning which assistants look things up live and change within weeks and which lean on what they learned long ago? Can they read sources, meaning look at an answer and say which review sites or guides it leaned on and why you are missing from them? Can they write a page with the answer first, specific details and honest limits? And can they measure, with a fixed question list, repeated runs and saved text? Ask them to run one question live on a call and walk you through all four. Fifteen minutes is enough to hear the difference.
When not to hire anybody yet
If you have never checked whether AI crawlers can reach your site, do that first. It is free, takes an afternoon and is the most common reason a business is invisible. If you have no baseline, start measuring before spending. And if your site is thin, with pages that repeat each other and nothing worth quoting, no answer-engine work fixes that, because the material does not exist yet. An agency worth hiring tells you this in the first conversation, not on the fourth invoice.
Twelve questions to ask on the first call
You do not need to be an expert to interview an agency. Take these to the call, or send them ahead and ask for written replies, because written answers commit the agency to wording.
Proof. (1) What questions do you measure, and who writes the list? Good: written with you, twenty to fifty questions, kept fixed. Weak: "our system generates it" or "it is proprietary." (2) Can I see a baseline from a past client, name removed, with the full answers? (3) How do you handle answers changing from run to run? Good: several runs per question, reported as a rate. (4) Show me a question where a client went from absent to named, and what you did.
The work. (5) Who writes the pages, and does a person review each one for accuracy about my products and policies? (6) What will you do outside my own website, by name? (7) What happens in the first thirty days? Good: an access check, an agreed question list and a baseline, then first pages. Weak: a strategy deck, with measurement starting in month three. (8) How do you decide what to write first? Good: from the gaps in real answers. Weak: from keyword research.
Terms. (9) What does the fee include, and what costs extra? (10) What do I own if I stop? Good: the pages, the question list and the measurement history are yours to export. (11) How long is the commitment, and how do I leave?
Limits. (12) What will you not promise? This is the best question on the list because it is hard to rehearse. A good agency answers quickly: no guaranteed placement, no single timeline for every assistant, no bulk-bought mentions, no revenue claims with false precision. A weak one pauses and says it is confident in its results.
Treat weak answers to questions 1, 2 and 10 as near deal-breakers, because they cover who writes the yardstick, whether a baseline exists and whether you keep the history. Ask every agency the same twelve in the same order and write the answers down in their own words, because memory flatters a fluent salesperson.
A worked example, labeled as invented
This exchange is made up to show how the same question lands with two agencies. A flooring store owner asks question 4. Agency one says: "A client in home improvement asked why a competitor was named for 'best flooring for a house with dogs.' The competitor had a page answering exactly that and the client had nothing on pets. We wrote one with specifics on the products they stock. About six weeks later, in repeated runs, the client began appearing in some answers. Not all. A second question on stairs did not move at all." Agency two says: "Our clients see great results across all their key questions." The first answer has a question, a gap, a fix, a time frame, a partial result and an honest miss. The second has nothing you could check.
Red flags, and when each one shows up
Warning signs appear at four moments: the pitch, the contract, the first month and the third month. Walking away at the pitch costs an hour. Walking away at month three costs money, time and often an argument, so learn to spot them early.
In the pitch. A guaranteed mention, since nobody controls which businesses an assistant names. A secret method that stays vague when you press, in a field where the basics can be described in one breath. One screenshot as proof. Rankings offered as proof of AI work.
In the method. Hidden text, instructions buried in a page or files sold as a switch that makes assistants name you. These tricks are dishonest and tend to be caught. Bought mentions and bulk placements, which assistants and the sites they read have become good at ignoring.
In the contract. Activity promised and outcome never mentioned, such as "twelve pages a month" with nothing you can hold them to. No baseline clause. A plan that never leaves your own website. A twelve-month lock-in for work whose results show slowly.
Once work has started. A question list that quietly changes, which rewrites the history. Traffic as the headline metric, when most recommendations produce no click. Reports with only good news, in a field where some questions always move down.
None of these proves bad faith alone. Two or three together are a clear signal. When you spot one, ask about it directly and kindly, negotiate the missing safeguard into writing, or, for a guarantee, a trick or a refusal to show proof, thank them and leave.
Read the contract and the price shape
Contracts in this field are young, and the weak ones are easy to spot once you know what to look for. Look for a baseline clause that requires a measurement before work starts. Look for an outcome the agency is willing to be measured against, not only a list of activities. Look for a short first term with an exit. And ask in writing what you keep if you leave. The measurement history is the most valuable piece and the one most often lost behind a login that closes when the contract ends.
The shape of the fee tells you what the agency thinks it is selling. A project price for the technical and editorial half is the cleanest arrangement, because the work is definable and you can tell whether it was done. A retainer suits the slow work but creates no pressure to show anything unless someone on your side reads the monthly report. Performance pricing tied to visibility needs care, because a question set the agency controls is a metric the agency controls. The arrangement that works most often is a defined project for the fast half, then a smaller retainer for depth once the agency has proved it does what it says. Our guide to AI SEO services and what they cost breaks down what drives the price and how to compare two quotes line by line.
Run a small test before a long commitment
You do not have to decide on the strength of a pitch. A small paid test, even a few weeks long, tells you more than any number of calls. Agree on ten to twenty questions that matter to your business. Run them across the assistants you care about and save the full answers. That is the baseline, and it will probably look discouraging, which is normal and useful.
Then choose two or three questions where the answer is plausibly fixable. Maybe the assistant names a competitor because it found a clear comparison page and you have none. Maybe it gets a fact about you wrong. Ask the agency what it would do about each, and have it do just those things. At the end, run the same questions again. You are looking for three things: did the agency explain what it did and why, were the answers saved properly, and did anything move in the direction it predicted? An agency that behaves like a careful scientist during the test will probably behave the same way during the engagement.
How RunOctopus does it
Everything above is how I would judge any agency, including ours. Here is how we do the work, in plain words.
RunOctopus is a done-for-you service. Our AI, Ollie, and our team research your business, then write, publish and link pages directly on your own website. Ollie does the research, drafting and linking, and people on our team plan, write, review and publish alongside it. It works with Shopify, WordPress, WooCommerce and Wix.
You do not have to approve each page first. There is nothing to install, upload or schedule, and pages go live on their own. That is a deliberate choice. The most common way this work fails is a good plan that nobody carries out, and an approval queue is where plans go to wait. We also report on where ChatGPT, Perplexity, Claude and Gemini name your business, so the measurement and the publishing sit in the same place.
We publish our case studies with the misses left in, measured from each client's own Search Console and from AI answer checks. A few of the numbers from them:
- All Angles Creatures, the reptile-feeder store where RunOctopus started, earned 5,447 Google clicks in the 28 days to October 10, 2026, up from 3,645 in the 28 days before. On two locked sets of AI questions it was the most-cited domain in every answer set, and ChatGPT named it on 4 of 8 generic buyer questions. It was absent on the generic isopod and springtail questions, and the case study says so.
- Florida Realtor Careers, a recruiting site, had 727 pages added between June 25 and September 22. 530 of them (73%) have appeared in Google, and monthly clicks went from 2 in June to 130 in September. On October 13 it showed for three of the four broad recruiting searches it was built to win, against none on June 26. The case study calls that win real and small.
- Sterman Multifamily, a Los Angeles apartment broker, earned 12 to 57 clicks a month for eight months, then reached 283 in August and 234 in September. A 238-page batch showed up fast and then thinned out, and that part is written up too.
If you want to see what Ollie would do for your site before talking to anyone, paste your address for the free site read. No account and no call.
A short way to decide
If you only remember five things, make them these. One, ask for a repeated measurement, not a screenshot. Two, ask how the hours split across measurement, access, content and outside sources. Three, ask who writes the pages and who reads the answers each month. Four, read the contract for a baseline, an exit and who owns the history. Five, run a small test first.
An agency that welcomes all five is probably worth talking to, whatever it charges. One that resists most of them is telling you how it expects to be judged.
An AI SEO agency should measure, fix access, publish answer-first pages and work on the sources assistants trust. Judge it on evidence you can inspect, and own the question set and the history.