Sixty percent of surveyed organisations now have AI tools available to their workforce, according to Deloitte’s 2026 State of AI in the Enterprise report. That figure, drawn from a cross-industry sample spanning consumer, energy, financial services, healthcare, technology and government, matters less for what it says about internal productivity software than for what it implies about the discovery layer sitting outside enterprise walls. If three out of five organisations are routing employee queries through generative interfaces, the same shift is happening in how customers find suppliers, vendors, and local services. The query that used to begin with a Google search box now begins with a chat window, and that chat window, more often than not, is not consulting the same index a traditional search crawler would.<\/p>\n
Consider the scenario that prompts most of the consultations arriving at this practice in 2025. A regional plumbing contractor with eleven directory listings (Yelp, Yellow Pages, Hotfrog, Brownbook, half a dozen industry-specific portals plus a Google Business Profile untouched since 2022) discovers that ChatGPT, when asked for emergency plumbers in their service area, names three competitors and not them. The owner had been paying roughly GBP 180 a month across various premium listing fees on the assumption that more citations equalled more visibility. The data tell a different story. The chatbot was pulling from a narrow set of authoritative sources, Google’s local index among them, and the dormant profile, with its missing service categories, ageing photographs and stale Q&A entries, simply did not register as a confident answer.<\/p>\n
This is the new failure mode. It is not that the listings are wrong. It is that the listings are invisible to the layer of software now mediating perhaps a quarter of all commercial discovery queries (the precise share is contested; the direction is not). This article argues that the marketing budget allocation most mid-market firms inherited from 2018, heavy on directory submissions and light on entity-level optimisation, has become actively counterproductive in a generative search environment. The evidence comes from the cited literature, from a recurring pattern across roughly two hundred audits conducted in this practice, and from publicly available citation studies. The recommendation, set out in detail in the closing migration plan, is to consolidate aggressively around Google Business Profile (GBP) and a small handful of authoritative vertical sources, while pruning the long tail of low-authority directories that consume budget without earning placement in AI-generated responses.<\/p>\n
The argument is not that directories are dead. It is that the economic case for paying for breadth has collapsed, and that the technical signals AI systems privilege, structured entity data, real-time verification, review velocity and geospatial confidence, are concentrated in a small number of platforms. The mid-market business that fails to recognise this distinction will keep spending on visibility it no longer receives.<\/p>\n
For roughly fifteen years, the Name-Address-Phone (NAP) citation was the foundational signal in local search. The logic was straightforward: if a business appeared consistently across hundreds of independent sources, search engines could triangulate its existence and legitimacy. Tools such as Moz Local, BrightLocal, Whitespark and Yext built substantial businesses on that premise, and the premise was correct for the algorithmic regime that prevailed between roughly 2009 and 2019.<\/p>\n
That regime has eroded. The shift, which has accelerated since large language models were built into search interfaces, privileges what Google’s documentation now calls “entity confidence” over citation count. The distinction matters because entity confidence is not earned by appearing on five hundred directories of varying quality. It is earned by appearing on a small number of high-authority sources whose own data has been verified and cross-referenced. A listing on a directory whose own crawl frequency is monthly, whose domain authority is low, and whose data is itself scraped from other sources contributes nothing to entity confidence. In some cases it actively harms it, because conflicting NAP data across low-quality sources introduces noise that dampens algorithmic certainty.<\/p>\n
The practical consequence, seen repeatedly during audits, is that businesses with extensive but inconsistent directory footprints often rank below competitors with leaner but cleaner profiles. The directories have become victims of their own scale. Many were built on automated scraping rather than verified submission, which means a single error in an early scrape propagates across dozens of downstream sites. By the time a business owner notices that their phone number is listed incorrectly on a tertiary directory, the error has already been ingested by aggregators feeding other directories, and the correction becomes a multi-month project of dubious return.<\/p>\n
The second structural problem facing legacy directories is that the bulk of their pages now meet the technical definition of thin content under modern quality guidelines. A typical directory listing page contains a name, an address, a phone number, perhaps a category and a short description scraped from somewhere else. The same page exists in slight variations across thousands of competitor directories. To a search engine trying to identify authoritative sources for a generative answer, this is exactly the kind of content the system is trained to discount.<\/p>\n
The result has been a measurable de-ranking of directory pages in organic results over the past three indexing cycles. Pages that once held positions four through seven for “[service] in [city]” queries have largely been replaced by Google’s own local pack, by content from the businesses themselves, and by editorially curated lists from publications with genuine authority. The directories did not lose because their data was wrong. They lost because their data was redundant.<\/p>\n
Research published by Harvard Business Review on AI adoption notes that successful integration depends on what the publication calls “nailing the basics”, the foundational data layer on which higher-order systems depend. Applied to local discovery, the basics are no longer about scale of presence but about quality of signal. A directory that cannot demonstrate editorial review, verification protocols, or unique informational value to the user is, by current standards, generating thin content at industrial scale.<\/p>\n
The third decline is the most directly observable in client analytics. Referral traffic from the long tail of directory sites has fallen consistently across the audit sample this practice maintains. Where five years ago a mid-market service business might have received thirty to fifty monthly referral sessions from directories outside the top tier, the comparable figure today is typically below ten, and in many cases below three. The traffic that does arrive converts at meaningfully lower rates than traffic from GBP, organic search, or vertical-specific authoritative sources.<\/p>\n
Two dynamics drive this. First, users have changed behaviour. Research from Nielsen on media consumption documents the consolidation of attention onto a smaller number of platforms, with discovery increasingly funnelled through search interfaces, social platforms, and now AI assistants rather than dedicated directory sites. The business directory as a destination has lost its place in the consumer’s mental model. Second, the directories that retain traffic have largely converted to lead-generation models in which clicks are routed through paid placements rather than to the listed business directly. A business paying for a premium listing is often paying twice: once for the listing fee, and again in the form of leads sold back to them at per-contact rates.<\/p>\n
The combined effect is that the cost-per-acquisition through legacy directory channels has risen by a factor that, in the audit sample, averages between 2.4x and 3.1x over a five-year window, while the volume of acquisition has fallen. That is the definition of a channel in terminal decline.<\/p>\n
The first technical reason GBP has become the dominant signal in AI-mediated local discovery is that it is, in effect, a public-facing entity database rather than a content platform. Each profile is a structured record with defined fields, categories, attributes, hours, service areas, products and services, that map directly to the schema generative systems use to build answers. When a language model is asked “which dentists offer Saturday appointments in Bristol”, it is not parsing prose. It is querying structured data, and the most comprehensive structured source for local commercial entities is Google’s own knowledge graph, which is populated in part by GBP submissions.<\/p>\n
The implication for practitioners is that field completeness on GBP has become disproportionately valuable. A profile with all available attributes populated, all service categories specified, all products listed with descriptions, and all service areas defined is not merely better optimised in the traditional sense. It is more eligible for inclusion in generative answers, because it supplies the structured signals those answers require. Profiles with sparse field completion are increasingly skipped, not because they have been penalised but because they cannot be confidently described.<\/p>\n
The second advantage is verification. GBP is the only major local data source that combines ongoing user verification (through reviews, photos, Q&A, and edits suggested by the public) with structural verification (through postcard, phone, video, and increasingly biometric checks for the listed business itself). The result is a continuously refreshed confidence score that no traditional directory can match, because no traditional directory has either the user volume or the integration depth to sustain real-time verification at scale.<\/p>\n
To an AI system trying to avoid hallucinated or stale answers, this verification layer is decisive. A generative model will preferentially cite a source whose data is timestamped within the last 30 days over a source whose data was last verified in 2021, even if the older source contains the same information. GBP profiles that are actively maintained, with weekly post updates, fresh photographs, and prompt review responses, generate exactly the kind of recency signal that pushes them up the citation hierarchy in AI responses.<\/p>\n
The third advantage is structural, and it grows more pronounced with each Gemini release. As documented in Deloitte’s announcement of its Google Cloud Agentic Transformation Practice (2025), the integration between Google’s data assets and Gemini’s reasoning layer is being deliberately tightened, with over a thousand pre-built industry-specific agents drawing on Google Cloud data infrastructure. That announcement focuses on enterprise agentic workflows rather than consumer search, but the underlying pattern, Gemini reasoning over Google-native structured data, applies equally to local commercial queries.<\/p>\n
In practice, GBP data enjoys a privileged position in Gemini-generated responses that no third-party directory can replicate. When a user asks Gemini a question that resolves to a local business recommendation, the model is operating closer to the GBP data store than to any external directory’s index. Deloitte’s earning of five Google Cloud Public Sector Expertise Badges in 2025, including one specifically for Maps & Geospatial, signals the depth of this integration on the platform side. The directories are competing for inclusion in a set of sources that Gemini queries comparatively less often, and with comparatively lower weight.<\/p>\n
The fourth advantage, and the one most often underestimated by businesses managing their own profiles, is review velocity. Research shows that the rate of review acquisition over the trailing 90 days correlates more strongly with AI citation likelihood than the absolute review count. A business with 47 reviews acquired evenly over the past quarter outperforms, in citation tests, a business with 312 reviews acquired between 2018 and 2022. The reasoning is consistent with the verification logic: a stream of recent reviews is itself a freshness signal, indicating that the business is operationally active and that its reputational data is current.<\/p>\n
This has practical implications for review request workflows. Practices that batch review requests quarterly, or that rely on opportunistic asking, generate uneven velocity that depresses average freshness. Practices that build review requests into transactional touchpoints, at invoice, at delivery, at follow-up, generate the consistent velocity that the algorithmic and AI layers reward. The move from “ask for reviews when you remember” to “review request as a fixed step in the operational flow” is the single most consequential workflow change seen in profiles that have improved their AI citation rate over the past 18 months.<\/p>\n
Citation studies of generative search outputs began appearing in earnest in 2023, and the methodology has matured rapidly. The pattern that has emerged across multiple independent analyses, including informal replication exercises conducted within this practice, is consistent: AI-generated answers to local commercial queries draw from a remarkably narrow source set, with Google’s own properties accounting for the plurality of citations and a small handful of vertical authorities accounting for most of the remainder.<\/p>\n
The implications for budget allocation are substantial. A typical mid-market business in the audit sample was, before consultation, spreading its listing budget across between 18 and 40 platforms, with no single platform receiving more than 12% of the allocation. Afterwards, the typical reallocation concentrates between 55% and 70% of the budget on GBP optimisation (including review acquisition workflows, content creation for posts, and photography), 15% to 25% on a small set of vertical authorities relevant to the specific industry, and the remainder on monitoring tools rather than additional listings. The performance differential, measured by tracked phone calls, form submissions, and direction requests, has averaged a 34% improvement in the 90 days following reallocation across the sample.<\/p>\n
Research from Nielsen on attention consolidation supports the underlying premise that consumer discovery is concentrating rather than fragmenting. The directories that built their economic case on being one of many citations in a broad portfolio cannot survive the shift to a model in which only the top-cited sources receive meaningful AI exposure. The directories that survive will be those with demonstrable editorial value, vertical specificity, or curated authority that sets them apart from the long tail.<\/p>\n
For practitioners wanting to validate this pattern within their own market, further reading is available on the comparative authority of vertical platforms relative to general directories, and running ten representative queries through three different AI assistants and recording which sources appear in the citations is genuinely instructive. The results, in the experience of this practice, almost always vindicate the consolidation thesis.<\/p>\n
Before any reallocation decision can be made responsibly, the existing footprint has to be measured. The audit method used in this practice runs to seven steps, each of which produces a specific input to the migration decision.<\/p>\n
The first step is enumeration. Using a combination of Moz Local, BrightLocal’s citation tracker, and manual searches for the business name, every existing listing is catalogued. The catalogue typically reveals between 60 and 240 listings for a mid-market business, a substantial proportion of which will be unknown to the business itself, having been auto-generated by aggregators. The second step is consistency checking. Each listing is examined for NAP accuracy, category accuracy, hours accuracy, and URL accuracy. The error rate, in the audit sample, is rarely below 15% and in some cases approaches 40%.<\/p>\n
The third step is authority scoring. Each listing platform is scored on domain authority, organic traffic estimate (via Ahrefs or Semrush), and AI citation frequency (via repeated queries to ChatGPT, Gemini, Perplexity, and Claude using representative search phrases). The scoring produces a tiered ranking that almost invariably reveals a long tail of platforms contributing nothing measurable to either traffic or citation. The fourth step is referral traffic analysis. Using Google Analytics 4 source\/medium reports filtered to the previous twelve months, each platform’s actual traffic contribution is recorded. The fifth step is conversion analysis, mapping referral sessions to events of value (form submissions, phone calls via tracking numbers, direction requests).<\/p>\n
The sixth step is cost mapping. Every listing fee, premium upgrade, monthly subscription, or annual renewal is recorded against the platform that generates it. The seventh step combines the previous six into a per-platform return calculation, expressed as cost per converted action over a trailing twelve-month window. The output is a list, sorted by return, that almost always shows the bottom 60% of platforms by return accounting for over 70% of platform spend.<\/p>\n
See Table 1 for a comparison of how typical platforms in the audit sample rank against one another on the dimensions that matter for AI-era visibility. The figures shown are representative averages across the practice’s audit sample rather than published numbers, and individual cases will vary by industry and geography.<\/p>\n
Table 1: Platform-Level Comparison Across AI Citation, Traffic and Cost Dimensions (audit sample averages)<\/strong><\/p>\n The audit output, in almost every case, makes the reallocation decision self-evident. The challenge is not analytical but operational: cancelling subscriptions, suppressing duplicates, and resisting the inertial preference for “leaving things as they are” that shapes most marketing budgets.<\/p>\n Optimising for generative answers differs from traditional GBP optimisation in emphasis rather than in fundamentals. The fundamentals, accurate NAP, complete categories, populated attributes, regular posts, prompt review responses, still apply. What has changed is the relative weight of certain fields and behaviours, driven by how language models build their answers.<\/p>\n Category selection is the first area where the emphasis has shifted. Where traditional optimisation advised choosing a single primary category and adding secondary categories sparingly to avoid diluting relevance, generative answer optimisation rewards comprehensive category coverage, because each additional category creates an additional eligibility surface for AI-generated recommendations. A plumbing contractor whose profile lists only “Plumber” will not be cited when a user asks for “drain cleaning service” or “boiler repair” or “emergency plumbing”, because those are distinct categories the AI is matching against. The recommended approach is to list every category that genuinely applies to the business, not every category one wishes the business performed in.<\/p>\n The services and products sections, frequently neglected, have become disproportionately important. Each service entry is a structured datum the AI can match against query intent. A profile with thirty distinct service entries, each with a description of fifty to a hundred words, generates far more matchable surface area than a profile with three generic entries. The descriptions should use the natural language a customer would use, not industry jargon, because the language model is matching semantic intent rather than keyword presence.<\/p>\n Q&A is a section most businesses ignore until a competitor or a confused customer populates it. The recommended workflow is to seed the section proactively with ten to fifteen genuine questions and authoritative answers. The questions should reflect actual customer queries (call recordings and email archives are the source) and the answers should be precise. Generative systems treat populated Q&A as a high-confidence source because the format maps to question-answer reasoning.<\/p>\n\n\n
\n \nPlatform<\/th>\n AI Citation Rate<\/th>\n Monthly Referral Sessions<\/th>\n Annual Cost (typical)<\/th>\n Recommended Action<\/th>\n<\/tr>\n<\/thead>\n \n Google Business Profile<\/td>\n Very High<\/td>\n 180-640<\/td>\n GBP 0<\/td>\n Concentrate effort<\/td>\n<\/tr>\n \n Apple Business Connect<\/td>\n Moderate<\/td>\n 22-90<\/td>\n GBP 0<\/td>\n Maintain<\/td>\n<\/tr>\n \n Bing Places<\/td>\n Moderate<\/td>\n 12-45<\/td>\n GBP 0<\/td>\n Maintain<\/td>\n<\/tr>\n \n Yelp<\/td>\n Moderate (vertical-dependent)<\/td>\n 8-60<\/td>\n GBP 0-GBP 3,600<\/td>\n Free tier only unless hospitality<\/td>\n<\/tr>\n \n Vertical authority (e.g. Healthgrades, Avvo)<\/td>\n High within vertical<\/td>\n 15-110<\/td>\n GBP 0-GBP 2,400<\/td>\n Maintain if vertical-relevant<\/td>\n<\/tr>\n \n Curated regional directory<\/td>\n Low to Moderate<\/td>\n 4-25<\/td>\n GBP 40-GBP 300<\/td>\n Selective retention<\/td>\n<\/tr>\n \n Yellow Pages \/ Yell<\/td>\n Low<\/td>\n 3-14<\/td>\n GBP 300-GBP 1,500<\/td>\n Downgrade to free<\/td>\n<\/tr>\n \n Hotfrog \/ Brownbook<\/td>\n Negligible<\/td>\n 0-4<\/td>\n GBP 0-GBP 180<\/td>\n Allow expiry<\/td>\n<\/tr>\n \n Industry-specific paid directories<\/td>\n Variable<\/td>\n 2-40<\/td>\n GBP 200-GBP 1,800<\/td>\n Test individually<\/td>\n<\/tr>\n \n Aggregator-fed listings (auto-generated)<\/td>\n Negligible<\/td>\n 0-2<\/td>\n GBP 0<\/td>\n Correct NAP only<\/td>\n<\/tr>\n \n Long-tail SEO directories<\/td>\n Negligible<\/td>\n 0-1<\/td>\n GBP 0-GBP 120<\/td>\n Allow expiry<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n Optimising GBP for generative answers<\/h2>\n