Read five agency pages selling AI search optimization and you will see the same menu five times: content optimization, entity building, technical implementation, AI visibility tracking, digital PR. The bullets are identical because the bullets are not the product. They are a translation layer between what the agency already sold last year and what buyers are searching for this year.
I sell this service, so read what follows knowing that. But I also publish my methodology and my price, which buys me the standing to do something the category will not: decode the menu. Here is what each line item actually is, where the real work happens, and the questions that tell you in one call whether a vendor, me included, is selling the work or the wrapper.
Decoding the standard menu
“Content optimization” usually means editing pages on your own website: restructuring for answer formats, adding FAQ blocks, tightening headings. There is a legitimate core here. Engines lift answers at the passage level, so writing passages an engine can quote cleanly, specific, verifiable, self-contained, is real craft, and the better shops do it at that level rather than sprinkling FAQ blocks. But it is also the single most oversold item on the menu, because your own site is the least-weighted input into how an AI engine describes you. An engine trusts what others say about you over what you say about yourself. A program that is mostly on-site editing is a program aimed at the weakest lever, and it is popular with vendors for an unflattering reason: it requires nobody’s cooperation and produces visible artifacts to put in the monthly report.
“Entity optimization” means schema markup, a Wikidata entry, consistent naming, sameAs links. This is a floor, not a lever. It helps an engine connect the dots it has already found; it does not create the dots. Entity clarity is a consistency problem in sources you do not control, and markup contributes the smallest share of it. Fair price for this work: a one-time project measured in hours.
“Technical implementation” is crawler access, page speed, clean structure. Genuinely necessary and genuinely commodity. It is plumbing. If the pipes are broken, fix them once; a monthly retainer line for ongoing “technical GEO” should make you ask what exactly recurs.
“AI visibility tracking” is measurement: running your buyers’ queries through the engines on a schedule and recording who gets cited. You need this. Just be clear that measurement is not movement. A tracker pointed at nothing reports nothing, precisely. If the proposal leads with the dashboard, the dashboard is the product.
“Digital PR” is the euphemism for the only item on the menu that touches the main lever: changing what credible third parties publish about you. Getting a stale comparison page corrected. Earning a mention in a roundup the engines actually cite. Getting an analyst or reviewer to use the right noun for your category. This is the source-layer work, it is slow, it depends on outreach and relationships, and it is chronically underweighted in proposals because it is hard to promise and harder to industrialize. In a real program it is the center of gravity, not a bullet near the bottom.
The one question that sorts vendors
Ask where the hours go: what share of the monthly work happens on your domain versus out in the sources that write your AI answer. Then listen.
A vendor whose program is mostly your own website is selling you repackaged SEO with new nouns. Nothing wrong with SEO, but it is not what you came to buy, and the tell is that everything in the scope can be completed without persuading a single third party of anything. The uncomfortable truth of this discipline is that the outcome lives on other people’s pages, so a credible program must spend most of its effort there: identifying which specific pages teach the engines your description, correcting the wrong ones, and earning the missing ones. On-site work earns its place as the supporting act, making your pages worth citing once somebody arrives to check.
What a real month looks like
For calibration, here is the shape of a working month in our program, which is public in full on the methodology page: research into the exact questions your buyers ask AI engines, six pieces written against the highest-intent gaps in those answers, outreach to the third-party pages that currently misdescribe you or omit you, landing-page work so the visits convert, and a measurement pass against a fixed query set, run repeatedly because single checks lie. Roughly speaking, the writing and the source outreach are the bulk of the hours. The markup, the tracking, the technical checks are real but small.
Any vendor’s month should decompose like that into named activities with named targets. “Ongoing optimization” is not an activity. It is a line item where hours go to be unaccountable.
Five questions to ask anyone, including us
Which specific sources currently write our AI answer, and which will you try to change first? A real practitioner can show you the method for finding out; here is ours. Better still, ask them to run two or three of your buyers’ actual questions through the engines live on the call and read the answers with you. A vendor who has never audited your sources, and cannot bear to look at them in front of you, is quoting blind.
What will you show us at 30, 60, 90 days? The honest answer includes leading indicators before citations: crawled and indexed, appearing on retrieval engines first, sourced corrections published. Then repeated citation runs against a fixed query set, and AI referral traffic in your own GA4, in your account, not a screenshot of theirs. The honest timeline has a flat stretch, and whoever names it before you ask is your shortlist.
What do you refuse to promise? The right answer includes guaranteed citations, specific placement in ChatGPT, and results on memory-based engines on any fixed schedule. Anyone who promises those is pricing your hope.
What happens to the work if we leave? Content you own, corrections that live on third-party pages, a documented query set: portable. Scores in a proprietary dashboard: not. The exit inventory tells you what the engagement actually produces. Ask about the contract shape in the same breath. A long lock-in is only defensible if the vendor can name the reason from the work itself, which exists: the timeline has a flat stretch, and a program judged at week six looks like a failure right before it is not. That is why my own initial term is six months, printed on the offer page. A lock-in with no stated reason is not protecting the program, it is protecting the vendor from your judgment.
Why should this not be done in-house? A confident vendor can answer honestly, because there is a real answer: most companies should start in-house and hire out when consistency becomes the constraint. A vendor who cannot name the situation where you should not hire them is not describing a service, they are describing a net.
What I got wrong
Early on I weighted my own engagements toward on-site work, for the same reason the industry does: clients can see it, approve it, and feel it happening. The months were full and the answers did not move, because the pages teaching the engines the wrong story were pages I had not touched and had not planned to. The program only started producing when the mix inverted, most of the hours going to source identification and outreach, with the client’s own site holding the proof the sources could point to. I now treat that mix as the definition of the service, which is why this post keeps returning to it: the wrapper is on your domain, the work is off it.
If the menu above matches what you want done, the version I run is on the offer page, price included. If you would rather run it yourself, the methodology is public and everything on this site teaches it. Both of those are better outcomes for you than buying five familiar bullets with a new name on them.