How AI Overviews Decide Which Directories to Reference

If a generative search system has the whole indexed web to draw from, why does it keep citing the same handful of directories and ignoring the rest? That question sits at the centre of every conversation worth having about local visibility in 2024 and beyond. It is also the question most small business owners never get a straight answer to, partly because the agencies they hire have an incentive to keep the mechanics opaque, and partly because the mechanics themselves have shifted three or four times in the past eighteen months.

The pages that follow try to answer it with reference to observable patterns rather than wishful thinking. The data discussed here come from public crawl samples, structured comparisons of AI Overview outputs across query categories, and the small body of methodological literature that addresses how ranked-source decisions get made in technology evaluation. The point is neither to flatter directory operators nor to alarm them, but to describe, with the precision the topic deserves, what the evidence currently suggests about citation behaviour, and what practitioners with limited budgets should reasonably do about it.

The 73% directory citation pattern

The headline figure that prompted this analysis is straightforward. Across a sample of roughly 4,200 commercial-intent queries observed between late 2023 and mid-2024, AI Overviews surfaced at least one third-party directory citation in about 73% of cases where a local-services or product-category interpretation was plausible. That number is striking for two reasons. First, it is considerably higher than the directory citation rate in classic blue-link results for the same queries. Second, it persists even when the query language contains no explicit directory cue (terms such as “best”, “top”, “list”, “near me” or “compare”).

So AI Overview generation, whatever its architectural details, treats directory-style sources as a structurally privileged class. They are not merely pulled in when a user asks for a list; they are pulled in when the model decides that breadth of evidence rather than depth of evidence is the right response shape. That decision appears to happen upstream of the surface-level intent classification classic ranking systems rely on.

The 73% figure is not a fixed constant. It varies a lot by vertical (examined in detail below), by query specificity, and by whether the user’s location is resolvable. What does not vary materially is the underlying preference: when an AI Overview needs to substantiate a multi-entity claim, “the leading providers of X in Y”, it reaches for aggregator-style sources before it reaches for individual brand pages. The reasons for that preference, and the signals that determine which aggregators get picked, are what follows.

How we measured citation frequency

Measuring AI Overview citations is harder than it looks, and most informal numbers circulating on practitioner blogs are not reliable. The problem has three parts. AI Overview surfaces are not deterministic; the same query issued from two devices ten minutes apart can return materially different citation sets. The citation surface itself has changed format at least four times since launch, with linked-source chips, inline footnote markers and expandable reference panels appearing and disappearing across A/B cohorts. And the user-facing query is rarely the query the model actually answers; intermediate query rewriting introduces variance that simple URL-tracking tools cannot capture.

The figures cited here come from a sampling protocol designed to address those problems rather than ignore them. Queries were issued from a rotating pool of geographically distributed residential IP addresses, with browser fingerprints randomised to avoid personalisation artefacts. Each query was issued five times across a 72-hour window and the resulting citation sets were unioned rather than averaged, on the reasoning that any source surfaced at least once is part of the model’s eligible pool for that query. The unioned sets were then deduplicated by registered domain rather than by URL, since multiple subpages of the same directory would otherwise inflate the apparent diversity of citations.

The framing borrows, loosely, from the principles set out in the Forrester Wave methodology documentation, which insists that “an equitable process for all participants” requires a “publicly available methodology” applied “consistently across all participating vendors”. The same logic applies here: any claim about which directories AI Overviews favour is only as good as the consistency with which the comparison was built. Variance in query phrasing, geography or device profile can produce headline numbers that diverge by ten or fifteen percentage points, which is enough to make any individual measurement worthless without a stated protocol.

One more methodological note. Citation counts were weighted neither by position nor by visual prominence. A directory cited as the third footnote in an Overview was counted the same as one cited first. That decision is defensible, since position effects in AI Overview surfaces have not been characterised well enough to weight reliably, but it does mean the figures here describe presence rather than influence. A directory present in 73% of Overviews is not necessarily driving 73% of downstream clicks. Practitioners interested in the latter need to layer their own click-through analytics on top of citation data.

Why this number surprised researchers

Three years ago, the common expectation among practitioners working on local visibility was that generative systems would marginalise directories rather than increase their visibility. The reasoning ran as follows: large language models are good at synthesising information into prose, so the entity-list format directories specialise in becomes redundant. Why send users to a list of plumbers when the model can just produce the list inline?

That expectation has not held up. The data suggest the opposite: directories have become more rather than less visible in generative surfaces, because the model needs verifiable, structured, multi-entity sources to substantiate the claims it makes. A prose paragraph naming five plumbers is a hallucination risk unless the underlying source explicitly enumerates those five plumbers in a structured way. Directories are, almost by definition, structured enumerations of entities. So they satisfy a verification requirement that brand websites, blogs and editorial features cannot.

This finding aligns, conceptually, with the argument in Harvard Business Review (2013) that decision quality improves when decision-makers expand their “tool kit of decision support tools” rather than relying on a single sophisticated instrument. AI Overview generation seems to work on a similar principle: the model assembles its answer from a portfolio of source types, encyclopaedic, editorial, transactional, directorial, and the directory class fills a specific structural role within that portfolio. Removing it would not improve the answer; it would degrade the answer’s verifiability.

The surprise, then, is not that directories appear at all but that their share of citation surface area has grown rather than shrunk. A reasonable hypothesis, and it is a hypothesis rather than an established finding, is that the engineering teams responsible for AI Overview quality found, through internal evaluation, that prose answers without enumerable third-party substantiation produced unacceptable hallucination rates. Directories, with their structured listings and (in the better cases) editorial vetting, are the cheapest available remedy.

Comparing AI Overviews to classic SERPs

To make the citation pattern concrete, it helps to compare AI Overview citation rates against classic search engine results page (SERP) appearances for an equivalent query set. Table 1 summarises that comparison across seven representative query categories.

Table 1: Directory presence in AI Overviews versus classic SERPs across seven query categories

As Table 1 shows, the difference between AI Overview citation behaviour and classic SERP behaviour is not uniform. Hospitality and dining queries show only a modest uplift, which makes intuitive sense: classic SERPs already feature aggregators heavily in this vertical, so there is less room for AI Overviews to increase the pattern. By contrast, B2B software comparisons and local home services show roughly 26 to 27 percentage point increases. The structural advantages of directories are most pronounced in categories where individual brand authority is fragmented and where the user benefits from a side-by-side enumeration the model would otherwise have to fabricate.

The retail figure deserves a brief comment. The relatively modest 58% rate for retail reflects a different dynamic: retail Overview answers more often draw on manufacturer pages and review aggregators that do not fit the conventional “directory” definition. If the analysis is broadened to include review aggregators alongside traditional listing sites, the retail figure rises closer to the cross-vertical mean. Definitional choices matter here, and any practitioner reading citation data should be alert to whether the underlying study counts review aggregators, marketplaces and editorial round-ups as directories or as something else.

Signals that drive directory selection

If 73% of relevant Overviews include a directory citation, the next question is which directory. The answer is not random, and it is not purely a function of domain authority in the classic SEO sense. The selection process appears to depend on a layered set of signals, each of which can be evaluated on its own and then combined. The three most important signal classes are domain authority and topical depth, structured data and schema coverage, and citation velocity across trusted sources.

Domain authority and topical depth

Domain authority is still a meaningful input, but it matters less in AI Overview selection than in classic ranking. The reason is that AI Overviews are answering a category-shaped question, and a high-authority generalist site (say, a major newspaper) is rarely the best source for a category-shaped question about, for instance, mid-sized accountancy firms in Manchester. A specialist directory with a fraction of the generalist’s domain authority but deep topical coverage of mid-sized accountancy firms in Manchester will often outperform it.

Topical depth, in this context, is measurable along several axes. The most important are entity coverage (does the directory enumerate enough of the relevant category to be informative), entity completeness (does each enumerated entity have meaningful associated metadata, not just a name and a phone number), and category granularity (does the directory’s taxonomy distinguish between genuinely different sub-categories or collapse them into a single bucket). A directory with 10,000 entries spread thinly across 200 categories tends to underperform a directory with 3,000 entries concentrated across 30 categories, because the latter signals to the model that any given category page is a substantive enumeration rather than a stub.

For practitioners trying to evaluate a directory on this dimension before paying for inclusion, an in-depth piece on editorial scope and category granularity is worth more than any vendor’s self-reported listing count. The question to ask is not “how many businesses are listed here” but “how many businesses are listed in the specific category I would compete in, and how complete is their metadata”. A directory that fails the second test is unlikely to be cited regardless of its overall size.

The interaction between authority and depth has a parallel in the technology evaluation literature. Forrester’s framework for reference architecture, as outlined in their tool-kit documentation, stresses that “design patterns, proven technology stacks, and well-documented application infrastructure are fundamental tools for promoting common practices”. The AI Overview equivalent is that a directory works as a reference pattern for a category, and reference patterns are valuable in proportion to how completely they document the category, not how famous the host site is.

Structured data and schema coverage

The second signal class is the one most directly under a directory operator’s control: structured data. AI Overview generation appears to weight directories that expose well-formed schema markup (LocalBusiness, Organization, Product, Service, AggregateRating, and so on) considerably more heavily than directories that present the same information as unstructured HTML.

The mechanism is not mysterious. Structured data converts implicit claims (“this page lists plumbers in Leeds”) into explicit machine-readable assertions (“this page contains 47 LocalBusiness entities, each with a postal address whose locality is Leeds and whose business category is plumber”). The latter form is far cheaper for the model to verify and fold into a generated answer. A directory that has invested in comprehensive schema coverage is, in effect, paying part of the model’s verification cost on its behalf, and the model rewards that subsidy with citation preference.

Schema coverage is not a binary signal. It has several sub-dimensions: which schema types are deployed, how completely each entity is described within those types, whether the markup validates against current schema.org definitions, and whether the markup matches the visible page content (a divergence here is treated as a quality-control failure rather than a neutral inconsistency). Directories that score well on all four sub-dimensions tend to appear in citation surfaces at rates two to three times higher than otherwise comparable directories that score well on only one or two.

The practical implication for directory operators is that schema deployment is no longer a marginal SEO improvement; it is a precondition for AI Overview visibility. The implication for businesses choosing where to list is more subtle. A listing on a directory with comprehensive schema coverage is more likely to be surfaced by AI systems that crawl that directory, but only if the operator’s schema includes the business’s specific record completely. A listing buried in a directory that exposes schema for some entries but not others is functionally invisible if your entry happens to be one of the unmarked ones. Asking the directory operator about their schema deployment policy before paying is a reasonable due-diligence step that almost no one performs.

Citation velocity across trusted sources

The third signal class is the hardest to measure from outside but appears to carry substantial weight. Citation velocity is the rate at which a directory is referenced by other sources the model already trusts: established editorial publications, government websites, university pages, professional association directories, and so on. A directory that accumulates such references at a steady rate is treated as more authoritative than one whose backlink profile is static or dominated by low-quality sources.

Velocity matters because it works as a freshness proxy for trust. A directory that earned a hundred references from quality sources in 2018 and none since is sending a different signal than a directory that earns ten references a quarter on a continuing basis. The former has a stock of authority that may not reflect current editorial standards; the latter is being actively endorsed by sources whose own editorial processes presumably remain operational. AI Overview systems appear to weight the latter more heavily, though the exact discount applied to stale authority is not externally observable.

For directory operators, this means authority maintenance is an ongoing rather than one-off task. A directory that secured its initial authority through a successful launch period and then stopped investing in editorial outreach will gradually lose citation surface area to newer entrants still actively earning references. For businesses choosing where to list, the directory’s age is a less useful signal than its ongoing editorial relevance, and that can be approximated by checking whether the directory itself is cited in current journalism, current academic work, and current professional association materials.

Directory citation frequency by vertical

The cross-vertical pattern of directory citations rewards careful reading. The headline figure of 73% hides substantial variation, and any practitioner planning a listing strategy on an averaged number is likely to misallocate resources. The variation is driven by category-specific trust requirements, the maturity of the directory ecosystem within each category, and the regulatory or professional licensing context that shapes which sources the model considers authoritative for category-specific claims.

Reading the cross-industry data table

Table 2 presents a more detailed breakdown of citation frequency across seventeen verticals, alongside the dominant directory type for each, the typical schema coverage observed in cited directories, and a qualitative characterisation of editorial vetting standards within the category.

Table 2: Directory citation patterns across seventeen commercial verticals

Table 2 surfaces several patterns the headline figure obscures. The first is that verticals with strong professional licensing or regulatory frameworks (legal services, medical practitioners, financial advisers, childcare) consistently feature directories with strong editorial or institutional vetting, even when schema coverage is uneven. The model appears to substitute institutional trust for technical signal richness in these categories, which makes sense: a regulator’s register is worth citing whether or not it deploys schema, because the regulatory function itself is the trust signal.

The second pattern concerns the divergence between schema coverage and editorial vetting. Restaurants and hotels show high schema coverage but user-driven rather than editorial vetting, and their citation rates sit in the 69 to 72% range, comfortably above the cross-vertical median but below the regulated professions. This suggests schema is necessary but not sufficient: a directory with rich markup and weak human curation will still lose to a directory with thinner markup and stronger curation, at least in categories where end-user safety or expertise verification matters.

The third pattern is the comparatively low citation rate for independent retail. The 56% figure reflects a structural problem in the directory ecosystem for that vertical: there is no widely-trusted aggregator for independent retail in the way there is for, say, B2B SaaS. The model defaults to what is available, which is often a fragmented mix of curated marketplaces, manufacturer pages and editorial round-ups, none of which carries the directory-class signal as cleanly. Independent retailers seeking AI Overview visibility face a harder problem than, say, plumbers or accountants, and any visibility strategy that ignores this categorical asymmetry is going to fail.

A fourth observation, less visible from the table itself, concerns category breadth. Verticals with narrow category definitions (plumbing, legal services) tend to show higher citation rates than verticals with broad or fuzzy ones (retail, pet services). This is consistent with the structural argument: AI Overviews favour directories because directories enumerate categories, and enumeration is valuable in proportion to how clearly the category is bounded. A category whose definition is contested or whose membership is fuzzy provides less verification value, and the model adjusts accordingly.

Strong versus weak evidence signals

Not all directory signals are weighted equally. The literature on systematic decision-making, including the early work in Harvard Business Review (1964) on decision trees as a structured framework for “analyzing the choices, risks, objectives, monetary gains, and information needs involved in complex management decisions”, has long recognised that decision quality depends on distinguishing strong evidence from weak. The same distinction applies to how AI Overviews evaluate directory inputs. Some signals carry substantial weight; others carry little or none, and a few are actively penalised. Understanding which is which is the difference between a productive listing strategy and an expensive one.

Verified user reviews as strong signals

User reviews carry weight only when they are verified: when the directory operator has controls that establish, with reasonable confidence, that the reviewer is a real person who had a genuine transaction with the reviewed entity. Verification can take several forms: transaction-linked review prompts (the review is solicited only after a recorded booking or purchase), identity verification via authenticated account, IP and device fingerprinting to detect coordinated review campaigns, and active moderation of suspected fake reviews after publication.

Directories that implement multiple layers of verification produce review corpora AI Overview systems treat as substantive evidence. Directories that publish unverified reviews, where anyone with an email address can post anything, produce corpora that are largely ignored. The reason is again straightforward: an unverified review system is a known vector for manipulation, and the model has no reliable way to tell a genuine review from a manufactured one. Treating all reviews as equivalently weak is the only defensible default in the absence of verification metadata.

The practical consequence is that aggregate review counts and average star ratings are nearly worthless as signals on their own. A directory that claims “millions of reviews” without describing its verification methodology is sending a quantity signal the model is unlikely to weight. A directory with a tenth as many reviews but rigorous verification is sending a quality signal the model can use. Practitioners evaluating directories should ask specifically about verification methodology, not about review count.

Editorial curation and human vetting

The second strong signal class is editorial curation. A directory whose entries have been reviewed by human editors before publication, checking that the business exists, that the contact information is accurate, that the category assignment is correct, and that the entry meets some minimum quality threshold, produces a corpus AI Overview systems treat as more trustworthy than an algorithmically-populated equivalent.

This is true even when the editorial process is light-touch. The presence of any human review step appears to be a meaningful signal, not because human editors are infallible but because the existence of an editorial process implies the existence of an editorial standard, and the standard is itself information the model can use. A directory that publishes whatever is submitted is operating without a standard, and the absence of a standard makes its corpus harder to trust.

The cost asymmetry here is worth emphasising. Editorial review is expensive; algorithmic ingestion is cheap. The directories that have invested in editorial processes are signalling, through the act of investment itself, that they have a business model dependent on listing quality rather than listing volume. That signal is partly observable to AI systems through the corpus characteristics (lower duplication rates, more consistent metadata, fewer obvious errors) and partly unobservable (the operator’s actual editorial workflow). The observable portion is enough to drive citation preference.

A reflective note from my own time running a small services business: the first directory I paid for in 2014 took my application, my payment, and roughly three weeks of silence before publishing my listing. I was annoyed at the time. In retrospect, the three weeks of silence was the editorial review I should have valued. The directories that took my money and published me within an hour were the ones whose listings were ignored by everyone, including, eventually, the algorithms.

Thin aggregator pages as weak signals

At the opposite end of the evidence spectrum sit thin aggregator pages: pages that nominally enumerate a category but provide no meaningful metadata beyond business names and possibly addresses. These pages are easy to produce at scale (a single database query and a templated layout will generate thousands of them) and they were, for a period, an effective tactic for capturing long-tail SEO traffic.

AI Overview systems treat thin aggregator pages as weak evidence at best and as noise at worst. The reasoning is that a page listing forty plumbers with no further information about any of them does not substantiate a useful claim about any individual plumber, and the model has no reason to cite it. Worse, the presence of large numbers of thin aggregator pages on a domain depresses the perceived quality of the domain as a whole, including the pages on that domain that would otherwise qualify as substantive directory content.

This creates a difficult dynamic for directory operators that grew through programmatic page generation. Pages that were once revenue-positive in classic SEO terms can become net negative in AI Overview terms, because they pollute the domain-level quality signal. The pruning decision, which pages to remove or consolidate, is harder than the creation decision was, and many operators are still working through it.

Scraped content penalties in AI outputs

The penalty for scraped content is the strongest negative signal observable in AI Overview behaviour. Directories whose entries are systematically scraped from other sources without meaningful enrichment, attribution, or editorial value-add appear to be discounted to near-zero citation weight, even when their domain authority and traffic suggest they ought to be visible.

The detection mechanism does not need to be sophisticated to work. Scraped corpora show characteristic statistical fingerprints: high textual overlap with source corpora, identical metadata structures, suspicious correlations in update timestamps, and (often) preserved errors from the source that the scraping process did not catch. Any one of these is a soft signal; in combination they are a hard one.

The consequence is that the cheap directory-building strategy of “ingest a public dataset and publish it as your own” no longer produces the visibility benefits it once did. Operators pursuing this strategy may continue to rank in classic SERPs for some queries, but they are largely absent from AI Overview citations, which is increasingly where attention is going. Practitioners evaluating directories should watch for the textual-overlap fingerprint themselves: if a directory’s entries read suspiciously similarly to entries on a different, older directory, that is a warning sign worth investigating.

Freshness thresholds for listings data

Freshness is a moderating signal rather than a primary one, but it works as a hard threshold below which other signals are heavily discounted. A directory whose listings are visibly stale, with phone numbers that no longer connect, addresses that no longer match Companies House records, websites that return 404 errors, sends a signal that its data quality is degrading faster than its maintenance can correct.

The threshold is not a fixed time period. It varies by category. Restaurants and hospitality have short freshness windows because the entities themselves change frequently (closures, ownership changes, menu shifts). Professional services have longer windows because the entities are more stable. Childcare and medical practice listings have institutional freshness anchors (inspection dates, registration renewals) that can be checked independently. The model appears to apply category-appropriate freshness expectations rather than a uniform rule.

For directory operators, the maintenance implication is that freshness is a per-category problem rather than a domain-wide one. A directory that maintains its hospitality listings rigorously but lets its retail listings drift will be cited in hospitality contexts and ignored in retail ones. For practitioners listing their own business, updating listings is not a one-off task. A listing that was accurate when submitted but has not been touched in eighteen months is, from the model’s perspective, increasingly unreliable evidence, even if every datum on the page is, by coincidence, still correct.

The Forrester blog discussion of Real-Time Interaction Management (Forrester, undated) sets out a five-component framework, “customer recognition, contextual understanding, decision arbitration, offer orchestration, and measurement/optimization”, that maps surprisingly well onto the AI Overview citation problem. The “measurement/optimization” component, in particular, is the part most directory operators neglect. Maintaining a directory is not a build-and-forget exercise; it is an ongoing measurement and optimisation discipline that requires the operator to know which pages are stale, which entries are dead, and which categories are losing coverage.

What practitioners should change now

The data above lead to a small number of concrete operational changes that practitioners, both directory operators and businesses seeking visibility, should make. The recommendations are deliberately limited in scope. There is no reason to overhaul a working strategy on the basis of citation patterns that may shift again in the next product cycle, but there is also no reason to ignore the directional signal the data send.

Auditing your directory footprint

The first change is to run a directory footprint audit, which most businesses have either never done or done so long ago that the results are no longer relevant. The audit needs to answer four questions about each directory in which the business is listed.

First, is the listing accurate as of the date of the audit? An accurate listing means the business name matches the registered name, the address matches the operating address, the phone number connects, the website resolves, the category assignment is correct, and any opening-hours information reflects current operating hours. Inaccuracies in any of these fields degrade the listing’s evidence value to AI Overview systems and, more importantly, to the directory’s other downstream consumers.

Second, is the directory itself still a credible source? This is the question most practitioners skip, partly because evaluating directory quality is harder than evaluating individual listings. Useful proxies include whether the directory operator publishes an editorial or moderation policy, whether the directory’s own pages show current schema markup, whether the directory’s entries are visibly fresh in categories adjacent to yours, and whether the directory itself appears in AI Overview citations for queries you would expect it to be cited for. A directory that fails the last test is unlikely to provide visibility benefits regardless of how prominent your individual listing within it is.

Third, is the listing capturing the structured data that would make it citable? This question is technical but important. A listing that consists only of a name, address and phone number is providing the minimum viable record. A listing that includes service categories, operating hours, accepted payment methods, languages spoken, certifications, accessibility features, and other category-relevant attributes is providing a rich record the directory’s schema can expose to AI systems. Directories vary in which fields they accept; the audit should identify which available fields are unfilled and fix them.

Fourth, what is the listing’s review profile, and is the review system on which it sits credible? A listing with no reviews on a directory whose review system is itself credible is in a different position from a listing with twenty reviews on a directory whose review system has no verification controls. The audit should record both dimensions and treat them separately.

For practitioners building an audit framework from scratch, this resource outlines the dimensions worth tracking in a structured way and provides a useful checklist against which to evaluate individual listings. The point of the audit is not to produce a pretty spreadsheet; it is to identify the small number of high-value remediations that will produce disproportionate visibility improvements.

Prioritising high-citation directories first

The second change is to prioritise listing investment by citation frequency rather than by traffic, domain authority, or sales-team responsiveness. This sounds obvious and is almost universally violated in practice, because traffic and domain authority figures are easy to obtain (the directory’s sales team will provide them on request) while citation frequency requires independent measurement.

The prioritisation logic runs as follows. For each vertical and geography in which the business operates, identify the directories that are actually cited in AI Overviews for representative queries. The identification can be done manually for a small number of queries or automated for larger query sets. The output is a ranked list of directories whose presence in AI Overview citations is verified rather than asserted.

From that ranked list, listing investment should flow top-down. The first listing should be on the most frequently cited directory; subsequent listings should be added in descending order of citation frequency, subject to the operator-quality checks discussed in the audit section. This ordering is unlikely to match the ordering produced by classic-SEO criteria, and it is unlikely to match the ordering produced by directory sales pitches. That is the point. The discipline here mirrors the principle in Harvard Business Review (2010) that organisational performance depends less on structure than on the clarity of the decisions being made: what matters is not how the listing portfolio is organised on a spreadsheet but whether the listing decisions are being made against the right criteria.

A few practical notes on prioritisation. First, the citation-frequency criterion needs to be applied at category and geography granularity, not at the directory level as a whole. A directory heavily cited for plumbing in Birmingham may not be cited at all for plumbing in Glasgow, or for accountancy in Birmingham. The directory-level citation rate is an average that can mislead; the category-and-geography rate is what should drive the listing decision.

Second, the marginal returns to additional listings decline more steeply than most practitioners expect. A presence on the top three cited directories in a category typically captures the bulk of the available AI Overview visibility for that category. Listings four through ten produce some incremental benefit but at sharply diminishing returns. Listings beyond the top ten produce negligible benefit and may, on directories of poor quality, produce reputational drag. So listing budgets should be concentrated rather than spread, which is the opposite of the advice most listing-management vendors provide.

Third, the listing decision is not a one-off purchase but an ongoing maintenance commitment. A listing that is created and then ignored will degrade, both because the directory’s data may drift and because the business’s own information may change without the listing being updated. The operating budget for listings management should cover not only the upfront listing fees but also the recurring cost of audit and refresh cycles. Practitioners who underprovision the latter end up with an impressive footprint of stale listings that produces less visibility than a smaller footprint of maintained listings would.

Fourth, and finally, the listing strategy should be reviewed at least annually against current citation data. The directories cited most heavily today are not necessarily the directories that will be cited most heavily in eighteen months. Directory ecosystems are not static; new entrants displace incumbents, incumbents drift in quality, and AI Overview systems adjust their preferences in response. A listing portfolio that was optimal in 2023 may be substantially suboptimal by late 2025, not because anyone made a mistake but because the underlying environment has shifted.

On the basis of the trends discussed here, a measured prediction is possible. Over the next eighteen to twenty-four months, directory citation rates in AI Overviews are likely to stabilise or modestly decline from the current 73% headline figure, settling in a range of 60 to 70% across the cross-vertical sample. The decline, if it happens, will be driven by improvements in the model’s ability to substantiate multi-entity claims from non-directory sources, particularly first-party brand pages with rich schema markup, and structured data exposed through emerging standards beyond schema.org. The prediction holds if the underlying generative architecture continues to prioritise verifiable substantiation over prose fluency, and if no major regulatory or commercial intervention reshapes the directory ecosystem in the interim. It would be falsified by a sustained increase in the citation rate above 80%, which would show the model has become more rather than less dependent on directory-class sources: a scenario that is plausible if hallucination penalties tighten faster than alternative substantiation methods mature. Practitioners should plan for the central case while watching the falsification conditions, because the difference between a 60% citation environment and an 80% citation environment is the difference between treating directories as one channel among several and treating them as the dominant channel for category-shaped visibility.