How do I get my business into Google's AI Overviews?
CANLAH AI is a Singapore SEO and GEO agency that treats AI Overviews as an outcome of ordinary search eligibility rather than a separate channel to be bought or gamed. The measurable work is crawlability, entity clarity, and passage-level answerability on pages Google already ranks.
Is there an official way to submit a business to Google's AI Overviews?
The short answer is no, and it is worth being blunt about that before anything else on this page. There is no queue, no form, no markup that asks, and no report that confirms. Understanding why that is structurally true, rather than a temporary gap Google will close later, is what makes the rest of the work legible.
There is no submission form and no separate index
Google publishes no mechanism for submitting a page to AI Overviews, no markup that requests inclusion, and no report confirming it happened. AI Overviews are generated from the same index and the same ranking systems that produce ordinary results; there is no parallel corpus to be admitted into. Anyone selling AI Overviews submission is selling something that does not exist. A page becomes citable by being eligible for Search — indexed, crawlable, previewable — and then by holding the most useful passage for the sub-question the query was decomposed into.
The sentence that is true and incomplete
Google's guidance says that nothing beyond ordinary search quality optimises for AI Overviews. That sentence is accurate about ranking mechanics and silent about failure modes. It describes what earns a citation; it does not enumerate what quietly disqualifies a page from being considered at all. Those disqualifications are documented, but they live under preview controls, rendering and indexing, on pages nobody is obliged to connect back to this question. The gap between the two is the only part of this problem an agency can honestly claim.
Why does most of this question's traffic land on Google's own documentation?
Being candid about search demand is unusual on a commercial page, so here it is up front: most of the traffic behind this question flows to Google's own documentation, and mostly it should. That fact shapes what a page like this one can usefully be.
The dominant intent is procedural
People typing this query want the button. Google Search Central answers procedural questions authoritatively, for free, and with the standing that comes from operating the system being asked about. Look at how this root is actually reformulated and most of the variants are hunting for official Google guidance rather than a vendor — Search Central and the developer documentation are the destinations. A commercial page competing head-on with the operator's own manual on how do I get in will usually lose, and probably deserves to.
Which makes this query a filter, not a lead source
Commercial intent inside this phrasing is thin. Treated honestly, the query qualifies rather than converts: readers who keep going after learning there is no button are the ones who suspect they have a real eligibility defect, and eligibility defects are engineering work with a testable outcome. Writing this page to beat Google's manual would produce a worse page than writing it for the minority who need the part the manual does not cover.
What the manual structurally cannot do
Documentation describes a system in general. It cannot tell a specific business which of its pages is excluded, why, in what order to repair them, or whether a repair worked. That is the difference between a specification and a diagnosis, not a shortcoming in the writing. Everything below assumes the manual has been read.
What part of AI Overviews can actually be engineered?
Splitting the problem into eligibility and selection makes the rest of it tractable, and it is the split most vendor decks avoid because it shrinks what they can sell. One side is deterministic, testable and permanently fixable. The other is probabilistic and, in most respects, not yours to control.
Eligibility is deterministic and testable
Eligibility asks whether a page can be used at all: is it indexed, does it render without a scripted fetch of the main content, does it permit snippets, is the answer inside crawlable HTML rather than an image or a panel that loads on click. Every one of those has a binary answer, a way to test it, and a fix. This is unglamorous technical SEO wearing a new label, and it is where most of the recoverable loss sits for a business that already ranks.
Selection is probabilistic and not steerable
Selection asks which of many eligible passages gets quoted for a decomposed sub-question, and that decision shifts between sessions, locations and phrasings. Under repeated sampling of a single prompt, three rounds returned three different vendor sets with no change made to any site in between; the measurement behind that is set out on our AI visibility page (our AI visibility check). Anyone offering deterministic control over this layer is describing a system that does not behave the way the observed system behaves. What improves is the odds: more eligible passages, more clearly attributable to a recognisable organisation.
Why does confusing eligibility with selection cost money?
The distinction is not academic. Nearly every disappointing engagement in this category can be traced back to buying one side while believing it was the other, and the failure looks different depending on which mistake was made.
Selling only selection
An agency that sells influence over which passage gets quoted has sold a promise it cannot keep, which is why such proposals are usually vague about verification. The tell is the absence of a defect list: eligibility work produces one automatically, because every finding is a specific URL in a specific broken condition. Selection work, honestly described, produces a sampling plan instead. A proposal with neither is selling neither.
Fixing only eligibility
The opposite failure is quieter. A business repairs its rendering and its preview directives, sees no citations, and concludes the channel does not work. Eligibility only makes a page usable; it does not make a mediocre page worth quoting. If the paragraph that answers the commercial question is a marketing sentence with no substance in it, restoring eligibility simply makes a weak passage eligible.
The order that works
Fix every eligibility defect first, because those are cheap, provable and durable. Then rewrite the passages that answer the questions worth money. Then treat selection as something to be sampled and reported rather than promised. That sequence also front-loads the parts a client can verify without taking anyone's word for it, which is the right way round for a relationship that has to survive its first quarter.
Which technical controls actually affect AI Overview eligibility?
These are the controls Google specifies precisely. Note the asymmetry before reading them: they mostly remove a page from consideration rather than promote it. That is exactly why they belong at the top of an audit — an exclusion nobody intended is the cheapest thing in this discipline to find and undo.
Snippet directives are the one fully specified lever
The nosnippet, max-snippet and data-nosnippet family controls whether Google may show text from a page in a preview, and restricting previews restricts the material available for a citation. These directives are frequently inherited from a legacy template, a cautious security review, or a publisher plugin nobody has opened in years. Auditing them costs almost nothing and occasionally explains an otherwise inexplicable absence. Removing a restriction restores eligibility; no directive exists that requests inclusion.
Google-Extended is a different question
Google-Extended governs whether a site's content may be used for grounding certain Gemini and Vertex products. It is commonly assumed to control AI Overviews, and that assumption does real damage when a nervous legal team blocks it expecting one outcome and gets another. Before changing any directive of this kind, write down which surface it is meant to affect and verify the mapping against current documentation. Guessing here removes visibility in a way that takes months to notice and longer to attribute.
Read the robots rules as a system, not a checklist
Crawl directives, preview directives and AI-training directives are separate systems that happen to share a file and a syntax. Teams review them as one block and change them as one block, which is how a decision about training data ends up suppressing a preview. Keep a written record of which rule exists for which reason.
What breaks eligibility most often in practice?
Sophisticated explanations get more attention than accurate ones. In practice, absence from generated answers is usually caused by something a competent engineer can find in an afternoon, and the candidates below account for most of it.
If the answer needs JavaScript, assume it is not there
Content injected after load, hidden behind a tab or accordion that fetches on click, or rendered only in a client-side route is unreliable material for a generated answer. The pragmatic test is to fetch the page without executing scripts and read what remains. If the paragraph that answers the question is missing from that output, the fix is server-rendered HTML, not more schema. This test takes one command and settles an argument that otherwise runs for weeks.
The boring layer decides most cases
Before anything sophisticated, confirm the page is indexable, canonicalised to itself, reachable without a query parameter, and not blocked by a rule inherited from a staging environment. A meaningful share of absence from generated answers is a crawl or canonical problem wearing a fashionable name. Check this layer first; it is the cheapest hour in any engagement and the one most often skipped because it is not interesting.
Test the page a machine sees, not the page you see
A browser carrying your session, your location and your extensions is not a useful witness. Fetch as an anonymous client, from a plausible location, with scripts off, and read the text that comes back. Most disputes about eligibility dissolve the moment somebody looks at that output instead of describing it from memory.
Which AI Overviews tactics are a waste of money?
The misconception specific to this topic is that AI Overviews reward a special format — that some arrangement of markup, headings or question phrasing acts as a key. Format helps a page be parsed once it is already in contention. It never causes inclusion.
Schema does not trigger an AI Overview
Structured data helps Google understand entities and relationships. It is not an inclusion signal for generated answers, and no property requests one. FAQPage markup in particular no longer earns a rich result for most sites, and it never earned a citation in a generated answer. Keep Organization and Article markup because they are cheap and unambiguous; drop anything bought specifically to trigger an overview.
Question-shaped headings with nothing underneath
Writing every heading as a question is not the tactic. The tactic is putting a self-contained answer immediately after the heading — one that stays true when lifted out of the page with no surrounding context. Pages fail this constantly: the heading asks, and the paragraph beneath opens with a pronoun or leans on something discussed above. Such a passage cannot be quoted, so it is not quoted.
Brand repetition inside the copy
Repeating a company name through the text in the hope that a language model absorbs it misreads how these systems work. Citation follows retrieval: the passage has to be retrieved for the sub-question before anything about it matters. Padding the text with the brand degrades the passage as an answer, which lowers the chance of retrieval, which is the opposite of the intended effect.
What else gets sold that does not move eligibility?
A few more purchases show up repeatedly in proposals written against this root. Each becomes easy to refuse once the eligibility-and-selection split is in hand.
Volume as a strategy
Publishing many thin pages around one query family produces near-duplicates that compete with each other and dilute the entity. It also reproduces the doorway pattern classic search guidance has always penalised, now with worse economics, because a generated answer needs only one good passage per sub-question. Fewer pages, each genuinely answering a distinct sub-question, outperform a content calendar built on permutations.
Paying for placement
There is no paid inclusion in AI Overviews. Advertising units render around them, which is a different product with a different buying process and a different measurement story. Treat any proposal that blurs the two as a warning about everything else in that proposal.
A monitoring dashboard sold as an intervention
A tool that checks a handful of prompts on a schedule is a measurement instrument, not a fix, and it is often priced as though it were both. Instruments are worth paying for when the sampling design behind them is sound. Ask what the tool does when a result changes: if the answer is that it draws a line on a chart, it is reporting, and should be priced as reporting.
What does a quotable passage actually look like?
Once eligibility is clean, the remaining craft is ordinary writing discipline that happens to matter more when a machine is doing the summarising. None of it requires private insight into how the model ranks, which is the test worth applying to any deliverable in this category.
Self-contained by construction
A quotable passage names its subject in the first clause, resolves every pronoun, carries no dependency on the sentence before it, and finishes within a couple of sentences. Test it by deleting the rest of the page and reading what is left. If the paragraph still makes sense to somebody who has never seen the site, it can be lifted. If it needs the heading to be intelligible, rewrite it until the heading is redundant.
One question, one passage
Generated answers are assembled from sub-questions, so a paragraph that answers two things at once is worse material than two paragraphs that each answer one. This is why long, hedged, everything-in-one-place writing performs badly here despite reading well to an editor. Split the hedge into its own passage and let the direct answer stand alone.
Keep the page a page
A page assembled purely from extractable fragments reads badly to humans and loses the coherence that makes a source worth trusting at all. Write something a knowledgeable reader would finish, then check that individual passages survive extraction. The order matters: coherent first, extractable second. Reversing it produces the kind of page that ranks briefly and convinces nobody.
Why does the organisation matter more than the URL?
Generated answers name organisations, not addresses. That shifts part of the work off the site entirely, onto the surfaces where an organisation is described by other people. This is the slowest work in the discipline and the part with the longest shelf life.
Make the organisation legible
Consistency across the site's own about page, business listings, professional directories and third-party mentions is a genuine asset: same legal name, same address format, same service description. Most Singapore businesses carry inconsistent variants across these surfaces after years of ad-hoc updates. Reconciling them is tedious, largely one-off, and among the few interventions whose effect persists across engine changes.
Be present where the citations already come from
Citations concentrate on a small set of sources per topic — trade press, marketplaces, forums, review sites, association directories. Identifying which sources an engine actually pulls for your queries, then being present and accurately described there, is more productive than another blog post. This is public-relations work with a search brief attached, and it is slow enough that it should be started early or not sold at all.
Other people's descriptions outrank your own adjectives
A sentence about a business written by somebody else carries weight that the same sentence on the business's own site does not. That is uncomfortable for anyone who wants full control of the wording, and it is why review platforms and directories punch above their traffic. Accuracy on those surfaces deserves more attention than the homepage headline usually gets.
How do you measure whether a business appears in AI Overviews?
Measurement is harder than optimisation here, and it is the part most reports get wrong. A single check proves almost nothing about a system that answers differently on consecutive runs, yet a single screenshot is what many audits are built on.
One check is not a measurement
A report built on one screenshot is reporting a single sample of a distribution and calling it a position. Sampling is the only honest approach: repeated runs, more than one engine, and a question set fixed in advance so results stay comparable after the site changes. Fixing the question set before the work starts also removes the temptation to quietly keep the questions that flattered the result.
A method has to be able to embarrass its operator
On a core commercial query in our own market, our self-audit result was named 0/5 — published rather than buried, with the working set out on our AI SEO page (our AI SEO guide). A method that cannot produce an unflattering finding about the agency running it is not a method. When choosing a vendor, asking to see a measurement where they did badly is the most informative question available.
What sampling cannot tell you
Presence rates do not measure revenue, and no current method attributes a purchase to a citation with confidence. Report presence as presence. Anyone converting citation counts into a projected pipeline figure is performing arithmetic on an assumption, and the assumption is the part worth arguing about rather than the arithmetic.
Is getting into Google's AI Overviews different in Singapore?
Mostly no, and saying so costs an agency nothing worth keeping. The mechanics of eligibility and passage quality are not regional. The genuine local differences are narrow, real, and worth naming precisely rather than inflating into a proprietary methodology.
The mechanics are not local
Eligibility, rendering, preview directives and passage quality behave the same here as anywhere else. In most cases there is no difference at all, and a vendor claiming a Singapore-specific method for AI Overviews is describing packaging rather than mechanics. What is local is the competitive set: fewer sites contest a given commercial query here, so an eligibility defect is more often the entire explanation for absence rather than one factor among many.
Where local really does change the answer
Query language mixes English with Malay, Mandarin and Singlish phrasing, and the same intent expressed differently can resolve to different source sets. Local business listings, .sg entity signals and regionally recognised directories carry more weight in establishing that an organisation genuinely exists and operates here. For service businesses, the sources an engine trusts on Singapore queries are more often local trade bodies and marketplaces than international publications.
On what this costs
Published Singapore retainer ranges for search and GEO work, and how to read the spread between them, are set out on our AEO and GEO page (our AEO versus GEO guide). The question to put to any vendor quoting inside that market is simple: which part of this fee is eligibility work with a defect list attached, and which part is sampling and reporting? A quote that cannot be split that way has not been thought through.
How long does it take, and what cannot be promised?
Boundary conditions belong on the page rather than in a sales call, because they are the part a buyer cannot verify later if it was only ever said out loud.
No guarantees, and the reason is structural
No agency controls the selection layer, so no agency can guarantee a citation, a position, or a date by which either arrives. What can be committed to is the eligibility work: defects found, defects fixed, verified by re-test. Timelines follow re-crawl and re-evaluation cycles that are not ours to schedule. A vendor guaranteeing an outcome here is either misunderstanding the system or counting on the buyer not measuring.
What a sensible first engagement covers
Audit eligibility across the pages that already rank; repair what is broken; rewrite the passages answering the commercial questions; reconcile the entity across listings; then establish a repeatable sample so the next report is comparable to this one. Anything that cannot be shown to a sceptical engineer — a defect list, a diff, a re-test — is not part of the deliverable.
How long does it take, and what cannot be promised?
The layer underneath this one
A related gap sits below the whole discussion: how a business is described to software agents that read catalogues and manifests rather than pages. In most stores that description is currently inherited from the platform's defaults rather than written by the brand, which is examined on our agentic SEO page (our agentic SEO guide). Worth knowing about now, and worth acting on after eligibility.
Where this page ends and documentation begins
For anything procedural about indexing, previews or structured data, Google Search Central remains the correct source and is better maintained than any agency summary. Read it first. What it cannot do is tell you which of your pages is broken, in what order to repair them, or whether the repair worked. That part is the engagement, and it is the only part worth paying for.