Something Inc.LoginSchedule a free consultation
GEO

AI Search Optimization Services: Scope By Surface

Google's two AI surfaces give the same answer and cite almost entirely different sources. Any AI search optimization services proposal that reports one visibility number is averaging systems that disagree.

GEOAI SEARCHNOV 2026

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.

DEFINITIONAI search optimization services are managed programmes that work to get a brand retrieved, named and cited inside AI-generated answers across Google's AI Mode and AI Overviews, ChatGPT, Perplexity and Claude, combining machine access work, content structure, off-site mention building and per-surface measurement.
TL;DR · 60 SECONDSThe evidence. The largest published comparison of Google's two AI surfaces is Ahrefs' analysis by Despina Gavoyannis, published December 15, 2025 on September 2025 US Brand Radar data, covering 540,000 query pairs for citations and 730,000 for content similarity. It found 13.7 percent citation overlap between AI Mode and AI Overviews, rising only to 16.3 percent across each system's top three citations, against 86 percent average semantic similarity in the answers themselves. An independent study by Victorious, published September 5, 2025 across 1,540 queries, measured 30 to 35 percent URL overlap and found 77 percent of unique domains appearing on one surface only, with no query producing identical citation lists. Our verdict: buy machine access and earned mentions first, insist that every report is broken out by surface with an error bar, refuse to pay a premium for schema as a citation lever, and treat any single composite visibility score as a reporting choice rather than a measurement. Both datasets are now roughly a year old, which is a real limitation and an argument for first-party testing rather than for ignoring them.
13.7%
citation overlap between Google's AI Mode and AI Overviews across 540,000 query pairs, in Ahrefs' analysis of September 2025 US data published December 15, 2025
86%
average semantic similarity between those same paired answers, with 89.7 percent of pairs scoring above 0.8. Same conclusions, different sources
77%
of unique cited domains appeared on only one of the two surfaces in Victorious' study of 1,540 queries, published September 5, 2025
45%
of AI Overview citations change between generations, per earlier Ahrefs research cited in the same analysis. The target moves while you measure it

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.

MEASUREAHREFS, 540,000 QUERY PAIRS, SEPT 2025 DATAVICTORIOUS, 1,540 QUERIES, SEPT 2025WHAT IT MEANS FOR SCOPE
URL citation overlap13.7 percent of citations were the same URL on both surfaces30 to 35 percent of AI Overview URLs also appeared in AI ModeBoth low. The magnitude depends on method, so ask which one a vendor is quoting
Top three citations16.3 percent overlap when restricted to each system's top threeNot broken outA prominent citation on one surface does not carry over to the other
Domain level overlapNot broken out77 percent of unique domains appeared on one surface only, 23 percent on bothDomain authority does not transfer between surfaces the way it does across Google's blue links
Answer similarity86 percent average semantic similarity, 89.7 percent of pairs above 0.8No query produced identical citation listsThe answer your buyer reads is stable. The credit for it is not
Wording overlapJaccard word overlap of 0.16, identical responses in 0.51 percent of casesNot measuredSame conclusion rewritten, which is why monitoring answer text alone misses the citation shift
Citations per queryNot broken outAbout 9 domains per query in AI Mode against 7.7 in AI OverviewsAI 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.

Machine access: server-rendered HTML, status codes, URL stability90%
Third-party mentions and earned media80%
Content substance: statistics, quotations, named sources75%
Per-surface measurement as a standing line item70%
Internal structure, headings and passage-level chunking65%
Schema markup as a citation lever30%
Composite AI visibility scores10%

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.

1Report every metric per surface, never blendedAI Mode, AI Overviews, ChatGPT, Perplexity and Claude get their own columns. Given 13.7 percent citation overlap between the two Google surfaces alone, a blended figure cannot be acted on: you will not know which surface moved or which workstream moved it.
2Sample repeatedly, then publish the spreadQuery each surface multiple times per measurement period and report the range alongside the central figure. With 45 percent of AI Overview citations changing between generations, a single pull per month produces a number whose error bar is wider than most of the changes being claimed.
3Separate mention from citation from groundingBeing named in an answer, being linked as a source, and being the document the model actually retrieved are three different outcomes with different causes. Collapsing them inflates good news. We set out the distinctions in how to measure grounding, citation and mention separately.
4Annotate the platform events on the chartCore updates, model swaps and reporting bugs belong on the timeline as vertical lines. Gemini 4 Hydrogen landing in Search during a core update is the kind of event that explains a month of movement better than any workstream in the scope does.
5Keep a first-party retrieval log you ownRecord the queries you care about, the answers returned and who was cited, on your own schedule and in your own store. It is cheap, it predates any vendor relationship, and it is the only evidence in this category that survives changing agencies.

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.

1AT THE PITCHAsk which surface they will move first, and why that one
THE MOVES
Require a named surface, not AI search in general, with a reason tied to your buyers
Ask what citation overlap they assume between surfaces and which study that number comes from
Reject the answer that all surfaces respond to the same work
DONE WHENYou have a named starting surface, a stated assumption about transfer between surfaces, and a source for it.
2SCOPING CALLAsk them to fetch your top pages without JavaScript, live
THE MOVES
Pick three of your own URLs during the call rather than letting them choose
Ask which templates they expect to fail before they run it
Require the per-template result as a deliverable, not a summary grade
DONE WHENYou know, page by page, what a retrieval crawler can and cannot read on your site, before any contract starts.
3IN THE PROPOSALAsk for the measurement methodology in writing
THE MOVES
How many samples per surface per period, and is the spread reported
Which of mention, citation and grounding each number represents
Whether platform events such as core updates, model releases and reporting bugs are annotated
DONE WHENThe methodology is specific enough that a second vendor could reproduce the numbers from it.
4BEFORE SIGNATUREAsk what they will refuse to promise
THE MOVES
Push for an explicit list of outcomes they will not commit to, in writing
Check that schema work is priced as hygiene rather than as a citation lever
Ask what result would make them tell you to stop spending
DONE WHENYou have a written boundary on the claims, which is the cheapest predictor of how the reporting will read in month six.
5IN THE CONTRACTOwn the measurement artifacts yourself
THE MOVES
Require the raw query and answer logs in your own storage, not a dashboard login
Agree that the methodology cannot change mid-engagement without a documented note
Keep the first-party retrieval log running independently of the engagement
DONE WHENChanging vendors costs you a relationship instead of your entire measurement history.

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.

DO THIS NEXTPick the ten queries that matter most to your pipeline and run each one five times, on the same day, on AI Mode and on AI Overviews separately. Record who gets cited each time. You will finish the afternoon with your own citation overlap figure, your own instability band, and a list of competitors that hold slots you do not. That exercise costs nothing and it will tell you more than most quarterly reports in this category. If you would rather hand the access diagnosis, the per-surface measurement and the mention building to a team that has run it across dozens of sites, that is what our ai search optimization services do inside a generative engine optimization engagement, and the no-JavaScript fetch on your own URLs is the first thing we will show you.

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.

What are AI search optimization services?Managed programmes that work to get a brand retrieved and cited inside AI-generated answers. Scope covers crawler access, content structure, third-party mention building and per-surface measurement. The output is merged changes and instrumented reporting, not a visibility grade.
How are they different from generative engine optimization?Generative engine optimization covers the whole problem, including presence in a model's training data. AI search optimization targets the live retrieval half, where the engine queries, fetches and cites. Retrieval is the half an engagement can realistically move in a quarter.
Do AI Mode and AI Overviews cite the same pages?Mostly no. Ahrefs measured 13.7 percent citation overlap across 540,000 query pairs of September 2025 US data, rising to 16.3 percent for top three citations. Victorious measured 30 to 35 percent URL overlap across 1,540 queries. Both figures are low.
Why do the two overlap studies disagree?They measured different things at different scales. Ahrefs compared exact URLs across 540,000 query pairs. Victorious compared cited URLs and domains across 1,540 queries and reported domain-level overlap separately, at 23 percent. Ask which figure a vendor is quoting.
Is a composite AI visibility score worth paying for?No. A score computed by an unpublished method cannot be audited, and blending surfaces that overlap 13.7 percent averages away the only actionable detail. Ask for per-surface figures with a sample count and a reported spread instead.
Does schema markup get you cited?The published evidence says no on its own. A February 2026 study reported a null result for schema presence predicting citation at p = .296, and Google states structured data is not required for AI Overviews or AI Mode. Treat schema as readability hygiene.
What actually correlates with being cited?Third-party mentions, on the strongest reported correlation, and substantive content. An analysis of roughly 75,000 brands put mentions at r = 0.664 against r = 0.218 for backlinks. The Princeton study measured up to 41 percent visibility gains from adding citations and quotations.
How long before results show?Plan two quarters before signal separates from noise, and insist on leading indicators meanwhile: retrieval access fixed per template, mentions placed, sample counts rising. With 45 percent of AI Overview citations changing between generations, early month-to-month movement is mostly variance.
Can I measure this in Search Console?Partly, and with care. The Generative AI performance report had a confirmed logging bug from November 9 to 11, 2026, and AI Mode link clicks are counted as ordinary clicks with no inline breakdown. First-party query logging covers what the platform reporting does not.
What should AI search optimization services cost?Price the remediation, not the audit. Access fixes usually need engineering time your team supplies, and mention building is the line item with real recurring cost. Ahrefs put the commercial term's average paid click at $19.00, which indicates what competitors pay for the same buyer.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.

JB
Josh BernsteinMANAGING PARTNER, SOMETHING INC.

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.

Keep reading

All insights →
GEOGenerative Engine Optimization Services: What To BuyGenerative engine optimization services usually arrive as a bundle: schema rollout, structured data, an entity page, directory submissions, some listicle placements, and a monthly citation report. The trouble is that the one line item most scope documents lead with is the one with the weakest published evidence behind it. A February 2026 study found no effect for schema presence on AI citations, and Google's own guidance says structured data is not required for AI Overviews or AI Mode. Meanwhile the strongest reported correlation with AI visibility is brand mentions on pages you do not own. Here is what a GEO scope should contain, in the order the evidence supports, what the market charges for it, and how to tell a real programme from an SEO retainer with new column headings.TYLER TRUFFI · 13 MINGEOAI Mode SEO: More Links, Still No Click DataOn October 15, 2026 Google announced what it called continued and substantial increases to the inline links inside AI Mode and AI Overviews answers, moved those links higher in responses more often, added image thumbnails to the Dive Deeper section, expanded timely news carousels to more countries and languages, and began testing a box at the top of an answer for the one or two official places on the web to complete a task. In the same week it confirmed that Search Console still will not break out clicks for AI surfaces. Barry Schwartz made the obvious point: links keep getting more prominent and the one metric publishers want keeps not arriving. Read those two facts together and the strategy changes. More inline links spread across more sources at a click rate Pew measured at roughly 1 percent is a redistribution, not a windfall, and the slot that actually carries value is the new official place box. AI Mode SEO now means competing for a two result slot you cannot measure, which makes first party instrumentation the only honest scoreboard.JOSH BERNSTEIN · 14 MINGEOGenerative Engine Optimization Agency: How to Vet OneGenerative engine optimization agency is a 2,200 search a month commercial term at keyword difficulty 15 and a $25.00 cost per click, the most expensive phrase in the agency category. Pull the result set and the proof layer is missing: a median URL Rating of 5 across nine ranking pages, an AI Overview citing a Reddit thread alongside a DR 39 agency page, and the two things a buyer most needs, how to choose and which agency type fits which business, scoring 38.7 and 33.9 out of 100 for coverage across that same set. Those are the worst covered topics in the category. This page is the evaluation layer the listicles skipped.TYLER TRUFFI · 16 MIN
Free consultation

Let us be the last SEO agency you ever work with

A 30 minute call and a free audit of your SEO and GEO position. You keep the findings either way.