Buyers shopping for AI search optimization services are shown one number: a visibility score, a share of voice, a citation count. The surfaces that number is computed from do not agree with each other. Google's AI Mode and AI Overviews, two systems from one company reading one index, answer the same question almost identically and credit almost entirely different sources.
That asymmetry is the most useful thing a buyer can know, and it is the thing almost no proposal mentions. If two surfaces inside Google disagree about who deserves the citation, then a programme scoped against AI search in general is scoped against nothing in particular, and a score that blends the surfaces reports the average of two systems that reached different conclusions about your brand.
The timing makes this sharper than it was a month ago. Google shipped Gemini 4 Hydrogen into Search in November 2026, which it called its biggest model upgrade to Search so far, and it landed in the middle of a core update whose ranking flux is still being tracked. The model generating the answers changed while the ranking system underneath it was moving.
What Are AI Search Optimization Services?
AI search optimization services are managed programmes that work to get a brand retrieved and cited inside AI-generated answers, covering crawler access, content structure, third-party mentions and measurement. The category is about three years old and its scope conventions were written by vendors rather than by evidence.
The commercial demand is real and it is unusually cheap to compete for. Ahrefs data recorded on October 7, 2026 put ai search optimization services at 600 US searches a month with a difficulty score of 8 and an average paid cost per click of $19.00. Across the full set of agency commercial terms checked that day, difficulty topped out at 28. Nobody has defended this ground properly, which is the only reason a term with that cost per click is still winnable.
The naming matters more than it looks. Generative engine optimization describes the whole problem, including being present in a model's weights from training. AI search describes the narrower and more tractable half: live retrieval, where the engine runs a query, fetches pages and cites them. The retrieval half is the half an engagement can actually move inside a quarter, and it is worth being explicit about which half you are buying when you compare a scope against a broader generative engine optimization services proposal.
Why AI Search Is Not One Surface
AI search is at least five distinct retrieval systems with different crawlers, different citation pools and different answer formats, and the two that share Google's index still disagree about sources most of the time. Scope and measurement both have to follow that split.
Two independent studies measured Google's two surfaces against each other: Ahrefs' comparison of AI Mode and AI Overviews and Victorious' citation study. They used different methods and got different magnitudes, which is exactly why both belong in front of a vendor. The agreement between them is the finding: overlap is low, and it is low at every level of granularity anybody has published.
| MEASURE | AHREFS, 540,000 QUERY PAIRS, SEPT 2025 DATA | VICTORIOUS, 1,540 QUERIES, SEPT 2025 | WHAT IT MEANS FOR SCOPE |
|---|---|---|---|
| URL citation overlap | 13.7 percent of citations were the same URL on both surfaces | 30 to 35 percent of AI Overview URLs also appeared in AI Mode | Both low. The magnitude depends on method, so ask which one a vendor is quoting |
| Top three citations | 16.3 percent overlap when restricted to each system's top three | Not broken out | A prominent citation on one surface does not carry over to the other |
| Domain level overlap | Not broken out | 77 percent of unique domains appeared on one surface only, 23 percent on both | Domain authority does not transfer between surfaces the way it does across Google's blue links |
| Answer similarity | 86 percent average semantic similarity, 89.7 percent of pairs above 0.8 | No query produced identical citation lists | The answer your buyer reads is stable. The credit for it is not |
| Wording overlap | Jaccard word overlap of 0.16, identical responses in 0.51 percent of cases | Not measured | Same conclusion rewritten, which is why monitoring answer text alone misses the citation shift |
| Citations per query | Not broken out | About 9 domains per query in AI Mode against 7.7 in AI Overviews | AI Mode has more slots, so it is usually the cheaper surface to enter first |
Read the answer similarity row against the citation overlap row and the implication is uncomfortable for the whole category. The engines have largely converged on what is true about a topic. They have not converged on who gets named for it. Content work that improves the answer can therefore succeed completely and still leave your citation share flat on the surface you happened to be measuring.
Be honest about the age of this evidence, because it cuts in both directions. Both studies rest on September 2025 data, and no comparable large-scale re-test has been published since. The direction is reliable. The exact percentages are dated, and the surfaces have changed underneath them: AI Mode started carrying more inline links in October 2026, which we covered when Google gave the surface more links without giving anyone click data.
“Two systems inside one company, reading one index, agree on the answer 86 percent of the time and on the source 14 percent of the time. A single visibility score across both is not a measurement. It is a rounding decision.”
What Should AI Search Optimization Services Include?
A credible scope orders its line items by published evidence strength: machine access first, earned third-party mentions second, content substance third, measurement as a standing cost, and schema last as cheap hygiene rather than a priced citation lever.
That ordering is not a house preference. It is what the record supports, and it is close to the reverse of how most proposals are sequenced, because the weakest line item is the easiest one to demonstrate in a slide.
Confidence a buyer should assign to each workstream, scored on whether a named reproducible mechanism connects the work to retrieval or citation. Illustrative scoring, based on the published evidence discussed in this section
The top bar is the one with real engineering behind it. Vercel's crawler analysis with MERJ, published December 17, 2024, found that none of the eight AI crawlers it tested executed JavaScript, and that 34.82 percent of ChatGPT crawler fetches landed on 404 pages against 8.22 percent for Googlebot. A page whose content is injected client-side is complete to Google and close to empty to a retrieval crawler, which is why access work is the only part of this category that behaves like ordinary engineering with a verifiable pass condition.
The second and third bars rest on correlation and on one controlled experiment. A widely reported analysis of roughly 75,000 brands put brand mentions against AI visibility at r = 0.664 versus r = 0.218 for backlinks, which is a correlation in a secondary write-up and not a causal finding. The Princeton generative engine optimization study presented at KDD in 2024 measured visibility gains of up to 40 percent from adding statistics and up to 41 percent from adding citations and quotations, which is a controlled result on a limited corpus. Treat both as direction, not as a conversion rate.
The bottom two bars are where money gets wasted. Kurt Fischman's February 2026 cross-platform study reported a null result for schema presence predicting AI citation, at an odds ratio of 0.678 and p = .296, and Google's own guidance states that structured data is not required for AI Overviews or AI Mode. Schema is useful readability work. It is not a growth tactic, and Profound's February 2026 research putting the share of AI citations explained by standard SEO ranking factors at 4 to 7 percent should set expectations for how much of this is reachable through classical optimization at all. Technical access and classical health still carry the Google side of the house, which is the argument we made in detail about scoping technical work for two crawler populations.
How Should AI Search Optimization Services Be Measured?
Measurement has to be per surface, repeated, and reported with a stated error bar, because the underlying citations are unstable by design and the platform-side reporting is both incomplete and occasionally wrong. One blended number hides all three problems at once.
Start from the instability. Earlier Ahrefs research found that 45 percent of AI Overview citations change between generations, meaning two identical queries minutes apart can return different sources. A single snapshot of your citation share is therefore a sample of one from a distribution nobody has characterised, and a month-over-month comparison of two such snapshots is noise presented as a trend. We built out the arithmetic of this in a framework for putting error bars on AI search measurement.
Then account for what the platforms do and do not tell you. Google confirmed a data logging bug in the Search Console Generative AI performance report between November 9 and November 11, 2026, which suppressed impressions for those three days. Any report covering that window and not mentioning it is reporting a platform defect as a performance change. Separately, Search Console counts an AI Mode link click as an ordinary click with no inline-link breakdown, so the surface that cites the most domains is the surface you can see the least about.
One consequence of all this is worth stating plainly, because it is the opposite of how the category sells itself. Good measurement here produces wider uncertainty, not narrower. A vendor whose confidence increases every quarter is either learning something nobody has published or smoothing the variance out of the report. The honest version of a monthly update includes the sentence that nothing moved outside the noise band, and most months on most surfaces, that is the true answer. For context on how far apart credible estimates of this market still are, the published numbers on AI search market share disagree with each other by multiples.
What To Ask Before You Buy AI Search Optimization Services
Five questions separate a programme built on the evidence from a retainer with new column headings. Ask them in this order, because the first answer calibrates how seriously to take the remaining four.
The shape of a bad proposal follows from those five. If the hours concentrate in schema rollout and directory submissions, the measurement is a vendor dashboard with an unpublished scoring method, and the reporting improves smoothly every month, you are buying a report rather than a result. The reverse shape, heavy on access diagnosis and earned mentions with measurement you own and uncertainty you can see, is where the economics work. Category buying patterns differ by sector too, and the picture is particularly sharp in B2B SaaS and software, where comparison and best-tool queries are the ones AI answers now intercept.
It is also worth saying what has not changed, because the 2026 framing tempts teams into abandoning work that still pays. Google remains the larger source of qualified traffic for almost every B2B site we work on, the November 2026 core update is still shaking out, and classical crawl, index and ranking health gates everything downstream of it. The AI surfaces added an audience with stricter requirements. They did not retire the old one.
Common Questions About AI Search Optimization Services
The questions below come up in every scoping conversation, answered short enough to forward to a finance lead who has no interest in search terminology.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.