What happens to a directory business when the search engine that sent it 60% of its traffic stops sending users altogether and starts answering questions directly from the listings instead?
That question sat at the centre of a six-month engagement that began in late 2023 and concluded in mid-2024. The client was a mid-sized vertical SaaS aggregator with roughly 14,000 indexed listing pages, a competent in-house developer, and a marketing director who had spent eighteen months watching impressions climb while sessions flatlined. What follows is a walkthrough of the audit, the twelve tactical interventions deployed, the sequencing decisions that mattered, and the measurable outcomes ninety days after rollout. Names are composited and figures rounded, but the methodology is reproducible.
The Client Brief That Started This
SaaS Directory With Stalled Traffic
The directory in question catalogued business software across roughly 240 sub-categories — payroll, CRM, document automation, that sort of territory. It earned revenue through a tiered listing model (free, enhanced, premium) and a small affiliate stream from outbound clicks to vendor sites. Monthly organic sessions in Q3 2023 sat at approximately 312,000, down from a peak of 487,000 eighteen months earlier. Impressions in Google Search Console, however, had risen by 22% over the same period. The classic decoupling: more visibility, less traffic.
The marketing director’s diagnosis was reasonable but incomplete. She suspected ranking volatility from the August and October 2023 core updates and had commissioned a content refresh that produced negligible lift. What the data actually suggested — and what a click-through analysis confirmed within the first week — was that AI-generated answers (Search Generative Experience at the time, plus increasingly assertive featured snippets) were absorbing the informational queries that had historically funnelled into category pages. The listings still ranked. Users no longer needed to click them.
Revenue had not collapsed in proportion to the traffic decline because premium listings were still being viewed by buyers further down the funnel. But the top-of-funnel drought was pushing CAC for paid acquisition past the point where the unit economics worked. The brief, agreed in October 2023, was deceptively simple: restore organic relevance in a search environment where the destination was no longer guaranteed to be a webpage.
Why AI Crawlers Changed Our Playbook
The conventional response to a traffic decline of this shape would have been a technical SEO audit, a content gap analysis against three or four named competitors, and a backlink campaign. None of those would have addressed the underlying problem, because the underlying problem was not ranking — it was retrievability for non-Google surfaces.
By late 2023, a meaningful share of category-level discovery traffic was being mediated by Perplexity, ChatGPT (with browsing enabled), Bing Copilot, and Google’s own SGE. These systems do not behave like classical crawlers. They retrieve, summarise, attribute (sometimes), and synthesise. A directory page optimised for ten-blue-link ranking is not necessarily optimised for being cited inside a generated answer. The two objectives overlap but diverge in important places — particularly around schema specificity, answer-extractability, and what one might call “citation density per paragraph.
The decision at this fork was whether to treat AI retrievability as an additive layer on top of conventional SEO, or as the primary objective with conventional SEO as a downstream beneficiary. The team chose the latter. That choice dictated everything that followed, including which tactics were prioritised, which were quietly deprecated, and how success would be measured. Research published by Deloitte Insights frames the broader pattern well: organisations are moving beyond reductive cost-cutting toward growth-oriented optimisation, and the directory business followed the same logic — fewer pages, better engineered, rather than more pages competing for the same shrinking pool of clicks.
Auditing the Existing Directory Structure
The audit ran for nineteen working days and used a stack that should be familiar to most practitioners: Screaming Frog (with custom extraction for schema fields), Ahrefs for backlink and SERP-feature data, Sitebulb for crawl visualisation, Google Search Console’s bulk export to BigQuery, and a manual review of approximately 400 listing pages stratified across category tiers. A separate workstream used the Common Crawl index to verify which pages were appearing in datasets likely to inform LLM training and retrieval.
Several findings shaped the eventual tactical plan. First, schema implementation was inconsistent — roughly 71% of listing pages carried Organization markup, but only 12% used SoftwareApplication with proper aggregateRating nesting, and almost none used FAQPage despite the page templates clearly containing question-and-answer content. Second, internal linking was heavily skewed toward category indexes; individual listings rarely linked sideways to related listings, which meant the semantic graph the directory presented to crawlers was star-shaped rather than meshed. Third, there was substantial content duplication at the boilerplate level — approximately 38% of every listing page consisted of identical template prose explaining what the category was, which any reasonable retrieval system would treat as low-signal noise.
The fourth and most consequential finding concerned what was missing rather than what was broken. Listings carried product descriptions, pricing tiers, and feature lists, but they did not answer the questions buyers were actually typing into AI search interfaces. A sample of 200 long-tail queries pulled from Search Console showed that 64% were question-formatted (“what is the cheapest payroll software for a 12-person team”, “does HubSpot integrate with Xero”), and the listing pages contained literal answers to perhaps 15% of them. The directory was, in effect, a catalogue presented as a catalogue, when the retrieval environment increasingly demanded a catalogue presented as a Q&A corpus.
An exhaustive review published in Artificial Intelligence Review (Springer, 2023) tracked roughly 540 metaheuristic algorithms and noted substantial similarities between methods marketed as novel — a useful caution when auditing any optimisation playbook. The lesson translated directly: most “new” SEO tactics circulating in vendor blog posts during late 2023 were rebranded versions of earlier techniques, and the audit team explicitly resisted adopting tactics that could not be validated against either log-file evidence or controlled before/after testing.
The audit closed with a prioritised defect register of 47 items, scored on three axes: estimated impact on AI retrievability, implementation cost in developer-days, and reversibility (a tactic that could be rolled back cheaply was preferred to an irreversible structural change, given the uncertainty of the environment). The twelve tactics that emerged were not the twelve highest-scoring items in isolation; they were the twelve that, sequenced together, would compound. That distinction matters and is the part most practitioner accounts skip over.
The Twelve Tactics We Deployed
Before drilling into the five tactics that warrant individual treatment below, it is worth listing all twelve in the order they were eventually deployed, because the sequencing tells its own story. The full set: (1) schema markup standardisation for listing entities, (2) semantic clustering of category pages, (3) answer-ready snippet blocks per listing, (4) canonical signals for AI retrieval, (5) programmatic page pruning, (6) entity-level disambiguation in page titles and H1s, (7) structured comparison blocks between adjacent listings, (8) freshness signals via verifiable update timestamps, (9) outbound citation hygiene to authoritative sources, (10) internal link mesh densification, (11) llms.txt and robots.txt recalibration for selective AI crawler access, and (12) sitemap segmentation by entity type rather than URL pattern.
Tactics six through twelve were largely mechanical once the first five were in place, and the sections below concentrate on the five that involved the most consequential decision-making.
Schema Markup for Listing Entities
Schema was the obvious starting point because it is cheap to implement, the impact is measurable within weeks, and AI retrieval systems demonstrably weight structured data more heavily than unstructured prose when assembling answers. The tactical question was not whether to add schema but which schema, at what granularity, and how strictly to validate.
The team standardised on SoftwareApplication as the primary type for every listing, nested inside an ItemList at the category level, with aggregateRating, offers (using PriceSpecification for tiered pricing), applicationCategory, operatingSystem where relevant, and a FAQPage block sibling to the main entity. Each listing also carried sameAs links pointing to the vendor’s verified social profiles, Crunchbase entry, and where applicable, Wikipedia or Wikidata identifiers. The Wikidata linkage was the single highest-effort element and the one with the most disproportionate downstream effect — entity disambiguation in AI retrieval depends heavily on knowledge-graph anchoring, and Wikidata IDs are the closest thing to a universal handle.
Validation ran through three layers: Google’s Rich Results Test for the formal schema spec, Schema.org’s own validator for completeness, and a custom script that verified every sameAs URL returned a 200 and contained matching entity references. Approximately 4% of listings failed the third check on first pass, usually because vendors had moved their Crunchbase pages or the directory had cached outdated social URLs. Cleaning these took longer than implementing the schema itself.
The decision not to use Product schema (which some competitors favoured) was deliberate. Product implies a transactional context that the directory could not honestly fulfil — it did not sell the software, only catalogued it. Misrepresenting the entity type to chase rich-result eligibility is the kind of short-term tactic that ages badly. SoftwareApplication told the truth about what the page contained.
Semantic Clustering of Category Pages
The directory’s category taxonomy had grown organically over six years and contained the artefacts you would expect: overlapping categories (“CRM” and “Sales CRM” and “Customer Relationship Management Software” all existed as separate URLs), thin categories with fewer than five listings, and a handful of categories that should have been sub-categories of broader parents. Before any AI-era tactics could have purchase, the taxonomy needed to make semantic sense to a retrieval system that would treat each category page as a node in a topical graph.
The clustering work used embeddings (OpenAI’s text-embedding-3-large, generated for every category page’s title, meta description, and top 500 words of body content) plotted with UMAP to identify natural clusters. The resulting visualisation showed seven clear super-clusters and twenty-three borderline cases where category boundaries were ambiguous. A human review — conducted by two members of the editorial team over four days — resolved each ambiguity by either merging, splitting, or re-parenting the affected pages.
The end result reduced the category count from 240 to 178, with 62 pages either consolidated into siblings or demoted to landing pages within a parent category. Every removed URL was 301-redirected to its semantic successor, and the redirects were monitored for thirty days to confirm equity transfer rather than dropoff. The cluster structure was then exposed explicitly in breadcrumb schema and in the sitemap segmentation, so retrieval systems received a coherent topical map rather than a flat list.
A comprehensive study on optimisation techniques in Artificial Intelligence Review (Springer, 2024) describes a shift toward biologically-inspired methods that produce rapid convergence on complex problems. The clustering work followed a similar logic — letting the natural structure of the data dictate the boundaries rather than imposing a top-down taxonomy that the directory had outgrown.
Answer-Ready Snippets Per Listing
This tactic addressed the audit finding that buyer questions were mostly unanswered on the listing pages themselves. The intervention was to add, on every listing, a structured Q&A block containing six to ten questions drawn from three sources: Search Console queries that landed on the page, “People Also Ask” data from Ahrefs, and a manual review of the vendor’s own support pages and review-site comments.
The questions were not generic. For a payroll software listing, the block might contain “Does [product] support contractor payments in Canada?”, “What is the minimum team size [product] economically supports?”, and “How long does [product] implementation typically take for a 50-employee company?” Each answer was constrained to between 40 and 80 words, written in declarative prose without marketing hedge, and marked up with Question and Answer entities inside a FAQPage schema block.
The constraint on answer length was deliberate. Retrieval systems extracting passages for generated answers tend to favour passages that can stand alone — too short and they lack context, too long and they get truncated unpredictably. The 40-to-80 word window emerged from testing against Perplexity and Bing Copilot during the audit phase: passages within that range were quoted intact roughly 73% of the time, while passages outside the range were either ignored or paraphrased into something the directory could not influence.
Production of these snippets at scale required tooling. The team built a workflow combining a custom GPT-4 prompt (for first-draft answers, fed with vendor-supplied factual data), an editorial review queue (every answer was checked by a human before publication), and a citation requirement (every factual claim had to link to a verifiable source — usually the vendor’s own documentation, sometimes a third-party review, never an unsourced assertion). Producing snippets for the top 2,400 listings took eleven weeks at a sustained pace of roughly 220 listings per week.
Canonical Signals for AI Retrieval
Canonicalisation in the classical SEO sense — telling search engines which version of duplicate content is authoritative — is well understood. What is less well understood is how AI retrieval systems decide which source to cite when multiple URLs contain near-identical claims. The behaviour is not deterministic, but the directory’s logs and citation tracking suggested that systems weighted three signals heavily: the freshness of the content, the density of original (non-boilerplate) text on the page, and the consistency of entity references across the page and its schema.
The team made several decisions here. First, every listing’s lastmod in the sitemap was tied to actual content changes, not to template updates or sidebar refreshes — which required reworking the CMS to distinguish between meaningful and incidental modifications. Second, the boilerplate template content was reduced from approximately 38% of page weight to under 12%, with the recovered space filled by listing-specific content. Third, every entity reference on a listing page (the product name, the vendor name, the category) used the same canonical string throughout, eliminating the variations that had crept in over years of editorial inconsistency.
A subtler decision concerned cross-domain canonicals. Some vendors had requested that their listing’s canonical tag point to their own website rather than to the directory page. This is a reasonable vendor request and a poor directory policy — it cedes authority on the directory’s own content to a third party and effectively asks search engines to ignore the listing entirely. The policy was reversed: every listing canonicalised to itself, and vendors who objected were offered a no-index option instead, which only three of approximately 600 affected vendors accepted.
Programmatic Page Pruning Decisions
Of all the twelve tactics, this was the most contentious internally. The directory had 14,000 indexed pages. The audit identified roughly 3,100 that were either thin (fewer than 200 unique words), stale (no meaningful update in over 24 months), or representing vendors who were no longer operating. The conventional view was that these pages still earned occasional traffic and removing them risked a net loss.
The counter-argument, which prevailed, was that retrieval systems penalise sites with high proportions of low-quality content not by demoting individual pages but by reducing the apparent authority of the entire domain. Keeping 3,100 weak pages to protect a few hundred trickle-traffic sessions was — to use an honest characterisation — a poor trade.
The pruning followed a three-tier protocol. Pages with zero organic sessions in the prior twelve months and no incoming backlinks were removed and 410’d (Gone, not 404, to signal intent). Pages with some traffic but no recoverable substance were merged into category-level content where the listed entity still made sense as a comparison reference. Pages with backlinks but no traffic were preserved with redirects to the most semantically appropriate live page. Approximately 2,200 URLs were ultimately removed, 600 were merged, and 300 were redirected.
The aftermath was instructive. Total indexed pages dropped to roughly 11,800. Average sessions per indexed page rose by 31% within sixty days. Impressions for the surviving category pages — which is the metric that matters most for retrieval visibility — climbed by 18% over the same window. The pruned pages did not represent traffic; they represented dilution, and the data confirmed it.
Sequencing the Rollout Over Six Weeks
Sequencing matters more than tactical selection in projects of this kind, because retrieval systems re-evaluate sites incrementally and a poorly ordered rollout can mask the impact of later tactics or, worse, attribute gains to the wrong intervention. The team allocated six weeks to deployment, with measurement windows extending through ninety days post-launch.
Week one was preparatory: backing up the existing site state in full, establishing baselines across every metric that would be tracked, and configuring the analytics pipeline to capture AI-referrer traffic separately from conventional organic. The latter required custom regex matching on referrer strings (Perplexity, ChatGPT, Claude, Bing Copilot, and Gemini all leave distinguishable signatures) and the use of a server-side log analyser, since client-side analytics misses a meaningful share of bot-mediated traffic. Without this baseline, attribution after the rollout would have been impossible.
Week two deployed the schema standardisation across all listings — a single coordinated push rather than a phased rollout, because schema is a passive signal and there was no benefit to gradual deployment. The CMS update took two developer-days; the validation sweep took five. By end of week two, 96% of listings carried correctly validated schema, with the residual 4% flagged for manual remediation.
Weeks three and four handled the taxonomy restructure and page pruning together. These tactics had to ship in the same window because removing pages without first having the redirect targets in place would have created a temporary equity sinkhole. The team used a Friday-evening deployment to minimise the user-facing disruption, monitored crawl error reports through Search Console and Bing Webmaster Tools over the following weekend, and had rollback procedures rehearsed (though not needed). For practitioners considering similar work, an in-depth piece on the topic of taxonomy restructuring at scale is worth consulting before committing to a single-window cutover, because the failure modes are unforgiving.
Week five was the answer-ready snippets — or rather, the first tranche of them, covering the top 600 listings ranked by historical traffic and revenue contribution. The remaining 1,800 high-priority listings rolled out across weeks six through ten, extending past the formal deployment window because content production at this quality level cannot be safely accelerated beyond a sustainable editorial pace. The decision to ship the highest-value listings first meant that measurement at the ninety-day mark would capture the snippets’ impact even if the full coverage was incomplete.
Week six closed with the smaller tactics — internal link mesh densification (executed via a script that identified semantically related listings using the same embeddings produced for the clustering work and inserted contextual links into listing prose, with editorial review of every insertion), llms.txt configuration (allowing GPTBot, ClaudeBot, PerplexityBot, and Google-Extended while monitoring whether allowing them produced measurable citation lift, which it did), and sitemap segmentation by entity type so that SoftwareApplication entities lived in a different sitemap from Article entities, simplifying both crawler hints and internal monitoring.
One sequencing decision deserves separate mention: the team deliberately did not pursue any link-building during the six-week deployment window. The reasoning was that a clean attribution window required eliminating confounding variables, and a backlink campaign run in parallel would have made it impossible to determine whether subsequent gains came from on-page work or from acquired authority. Link-building resumed in week ten, after the measurement baseline had stabilised.
Measuring Lift in AI Citations and Clicks
Measurement was structured around four metric families, each tracked daily and reviewed weekly. The first family was conventional organic — sessions, impressions, average position, click-through rate from Search Console — providing continuity with the pre-engagement baseline. The second was AI-referrer traffic, captured via the regex-matched referrer pipeline established in week one. The third was citation appearances in generated answers, measured by querying a panel of 480 representative buyer questions across Perplexity, ChatGPT, Bing Copilot, and Gemini on a weekly cadence and recording whether the directory was cited, paraphrased, or absent. The fourth was business outcomes — premium listing inquiries, affiliate clicks to vendor sites, and category-page conversion rate.
The citation panel deserves elaboration because it is the metric most practitioners get wrong. Querying AI systems is not deterministic — the same prompt produces different answers across runs, days, and accounts. The panel was queried five times per question per system per week (so 480 questions × 4 systems × 5 runs = 9,600 queries weekly), with citation rates aggregated to smooth the variance. Even with this volume, week-to-week noise was substantial, and any single week’s reading was treated as indicative rather than conclusive. Three-week rolling averages were the unit of analysis.
Findings from Logistics Systems: Design and Optimization (Springer) on the importance of measuring decision linkages — that is, how upstream decisions cascade through downstream metrics — informed the analytic frame. The team explicitly mapped each tactic to the metric it should most plausibly affect, and resisted the temptation to attribute every gain to the most recent intervention.
Numbers After Ninety Days
At the ninety-day post-launch mark (measured from the end of week six, so essentially mid-March 2024), the headline figures were as follows. Conventional organic sessions had risen from the pre-engagement baseline of 312,000 monthly to 389,000 monthly, a 24.7% increase. Impressions had risen 9% — a smaller proportional gain than sessions, indicating that click-through rate had improved on the pages that mattered. Average position across tracked terms improved by 1.4 places, which is meaningful but not dramatic.
The more interesting numbers came from AI-referrer traffic. Pre-engagement, the directory recorded approximately 1,800 monthly sessions attributable to AI referrers (mostly Perplexity, then ChatGPT, with Bing Copilot a distant third). At ninety days, that figure had risen to approximately 11,400 monthly — a 6.3x increase. The absolute number was still small relative to total traffic (under 3% of sessions), but the trajectory was clear and the conversion rate of AI-referred sessions was, notably, 2.1x higher than conventional organic, suggesting that users arriving via AI mediation arrived with stronger intent.
Citation appearance rates in the panel rose from a baseline of 8.2% (the directory was cited in roughly one in twelve generated answers across the panel) to 21.7% at ninety days. Perplexity drove the largest gain (the directory’s citation rate there went from 11% to 34%), with ChatGPT showing more modest improvement (6% to 14%) and Bing Copilot landing in between. Gemini was inconsistent throughout and did not show a stable pattern within the measurement window.
Business outcomes followed with a lag, as one would expect. Premium listing inquiries rose 18% month-over-month in months two and three post-launch. Affiliate click-through to vendor sites rose 27%, partly because the answer-ready snippets often included a natural call-to-action toward the vendor’s pricing page and partly because the higher-intent AI-referred traffic converted at the rates noted above. Total revenue attributable to organic discovery (a calculated figure, not directly observed) rose by approximately 22% over the ninety-day window.
Two cautions are warranted about these numbers. First, the directory operated in a vertical (business software) where AI-mediated discovery has matured faster than in many other categories — buyers in this space are technically literate and adopted Perplexity and ChatGPT as research tools earlier than the general population. Results in less AI-native categories will likely show smaller lifts, at least in the near term. Second, ninety days is a short window for a structural intervention of this scale, and some of the gains may compound further while others may attenuate as competitors adopt similar tactics.
Principles That Transfer to Any Directory
Setting aside the specific tactical implementations, several principles emerged from this engagement that appear to transfer across directory types, scales, and verticals. They are worth stating explicitly because they should outlast the specific tactics, which will themselves age as retrieval technology evolves.
The first principle is that retrievability is not a synonym for ranking. A page can rank well in classical search and still be invisible to AI retrieval, and vice versa. The two objectives share a great deal of underlying technical practice (clean schema, fast load, clear information architecture) but diverge in important ways — particularly around answer-extractability, entity disambiguation, and the granularity of structured data. Treating them as a single problem produces tactical decisions that under-serve both. Directory operators who continue to optimise solely for ten-blue-link ranking are, at best, leaving citation traffic on the table and, at worst, accelerating their irrelevance as the share of mediated discovery grows.
The second principle is that less content, better engineered, beats more content, weakly engineered. The pruning of 2,200 pages was the single most controversial decision in the engagement and arguably the single most impactful. The instinct to preserve every page that has ever earned a session is understandable but increasingly misaligned with how retrieval systems evaluate domain quality. As Deloitte Insights observes in its framing of optimisation as growth-oriented rather than reductive, the goal is not to cut for the sake of cutting but to remove what dilutes so that what remains can perform.
The third principle is that schema is necessary but not sufficient. Every directory operator with technical competence has implemented schema by now. The differentiator is no longer presence of schema but the granularity, accuracy, and completeness of the entity references — particularly the linkage to external knowledge graphs (Wikidata, Crunchbase, vendor-verified profiles). Schema without entity anchoring tells retrieval systems what kind of thing the page describes; schema with entity anchoring tells them which specific thing, and the difference determines whether the page is cited or generically paraphrased.
The fourth principle is that buyer questions, not category labels, are the organising unit of useful content. The audit’s finding that 64% of long-tail queries were question-formatted and only 15% were answered on the listing pages was not unique to this client. Most directories were built around taxonomies that matched how their internal teams thought about the content rather than how buyers expressed their needs. Closing that gap — by adding answer-ready snippets to every listing — is a high-yield, structurally simple intervention that almost any directory can execute, and the work conducted by the comprehensive review of metaheuristic methods in Artificial Intelligence Review (Springer, 2024) reinforces the broader point that the right unit of analysis is often the one buyers (or in their case, problem instances) actually present, not the one practitioners are accustomed to using.
The fifth principle is that measurement has to evolve to match the environment. A directory operator who in 2024 still tracks only sessions, average position, and conversion rate is missing the metrics that matter most for the next phase. Citation appearance rate, AI-referrer traffic, and the conversion-rate differential between mediated and direct organic traffic are now first-order metrics, and the tooling to track them — while still maturing — is good enough to support meaningful decision-making. Operators waiting for vendor-supplied dashboards to standardise these metrics will spend the next two years optimising for yesterday’s surfaces.
The sixth and final principle, more philosophical than tactical: the directory category is being unbundled by AI mediation, and the directories that survive will not be the ones that resist the unbundling but the ones that position themselves as the canonical structured-data source for their vertical. A directory whose entries are the most accurate, most comprehensive, and most retrievable representation of the entities in its space becomes infrastructure that AI systems cite rather than commodity content they paraphrase. That is a strategic position worth pursuing even where the short-term tactical investments are difficult to justify on a single-quarter ROI basis.
Adapting the Approach to Different Constraints
Working With a Five-Figure Budget
The engagement described above ran on a six-figure budget. Most directory operators do not have that. The realistic question is: which of the twelve tactics deliver disproportionate value at a fraction of the cost, and which can be safely deferred?
The answer, on the evidence of subsequent smaller engagements the practice has run, is that schema standardisation, page pruning, and a focused subset of answer-ready snippets account for roughly 70% of the gains achievable from the full programme. A directory operator with $30,000 to $60,000 to spend over a quarter can credibly tackle these three. The taxonomy restructuring is more expensive in human-judgement terms (it requires editorial knowledge that does not fully delegate to scripts) but can be done in phases, starting with the worst-offending overlapping categories and extending over multiple quarters as budget permits.
What does not survive a budget cut is the citation-tracking panel — at the volumes required for stable readings, it is operationally expensive (the fully-loaded cost of running 9,600 weekly queries plus the analyst time to interpret them lands around $4,000 per month). Smaller operators can substitute a manual sampling protocol: pick 50 representative questions, query them across two or three AI systems weekly, and accept that the readings will be noisier. The methodology degrades gracefully; the principle of measuring citation appearance does not require enterprise-scale instrumentation to be useful.
The internal-link mesh densification work also scales down well, because the embeddings generation costs are negligible (a few dollars in API fees for a directory of this size) and the script that inserts links can be written in a day. The constraint is editorial review — every inserted link should be checked by a human, which scales linearly with content volume. For directories under 2,000 listings, the manual review is tractable for a single editor over a few weeks. Above that, it requires either tooling investment or accepting some quality risk.
Local Service Directories Versus SaaS
The tactics described above were developed for a SaaS directory and several adjustments are necessary when applying them to local service directories — plumbers, dentists, accountants, the kinds of entities that appear in classical Yellow Pages successors and modern equivalents.
The schema differences are the most obvious. Local service entities use LocalBusiness (and its many specialisations — Plumber, Dentist, AccountingService) rather than SoftwareApplication, and the important fields shift from features and pricing tiers to address, geo, openingHoursSpecification, and areaServed. The aggregateRating field carries even more weight than it does for SaaS, because local-intent queries weight reputation signals heavily.
The buyer questions are different in shape. SaaS questions tend to be feature-comparative (“does X integrate with Y?”). Local service questions are more situational (“is X open on Sundays?”, “does X serve [specific neighbourhood]?”, “how much does X charge for [specific service]?”). Answer-ready snippets in a local context need to address these, and the snippets must be kept current — a SaaS pricing page that is six months out of date is annoying but tolerable; a plumber’s emergency-service hours that are six months out of date is a customer-experience failure that the directory will be blamed for.
The citation environment is also different. AI systems mediate a smaller share of local discovery than they do of SaaS research, partly because Google Maps and the local pack still dominate local-intent queries and partly because users searching for a plumber are typically further down the funnel than users researching software. The implication is that local directories should weight their tactical investments more heavily toward conventional local SEO (Google Business Profile management, citation consistency across the NAP ecosystem, review acquisition) and treat AI-retrievability tactics as a smaller, supplementary investment rather than the primary objective. As detailed in research published in a 2024 review of online discovery patterns, local-intent search behaviour has been slower to migrate to AI-mediated surfaces than B2B research-intent search, and the tactical balance should reflect that.
Two-Week Emergency Timelines
Sometimes the brief is not “improve our directory over six months” but “we got hit by an algorithm update last week and need to stop the bleeding”. The twelve-tactic programme does not compress to two weeks, but a defensible emergency triage does exist.
The emergency protocol prioritises three actions, in order. First, identify the pages that have lost the most traffic in the affected window using Search Console comparison reports. If these are concentrated in a particular category or topic, the issue is likely thematic; if they are spread across the site, the issue is structural. The diagnosis dictates everything that follows.
Second, audit those pages for the most common AI-era failures: missing or invalid schema, boilerplate dilution, missing answer-ready content, and weak entity references. Most algorithm-update casualties this practice has examined since mid-2023 fail at least two of these checks. Fixing them does not always restore the lost traffic — sometimes the loss is permanent and reflects a genuine shift in how the page should be valued — but it stops further deterioration and provides a foundation for recovery.
Third, do not panic-publish new content during the emergency window. The instinct in algorithm crises is to ship more content as a demonstration of activity, and it almost always backfires. New thin content compounds the problem the algorithm was already penalising. The two-week window is for repair, not expansion.
The fourth thing one might do — and the temptation is real — is to start removing pages aggressively in the hope of triggering a re-evaluation. This usually does more harm than good in a two-week window, because the redirect chains and equity-transfer dynamics need months to settle and an emergency context does not provide that runway. Page pruning is a tactic for stable conditions, not crisis ones.
Solo Operators Without Dev Support
A meaningful share of directory operators are solo or near-solo — a founder plus a part-time editor, perhaps, with no in-house developer and limited budget for outsourced technical work. The twelve-tactic programme assumes developer support and fails gracefully without it.
The tactics that survive the no-dev constraint are those executable through CMS interfaces or through plugins. Schema markup, on most modern CMS platforms (WordPress with Yoast or Rank Math, Webflow with native schema fields, Ghost with custom code injection), can be implemented without developer work, though the granularity will be lower than a custom implementation achieves. Answer-ready snippets are pure content work and require no development at all. Page pruning is a content-management task. Internal linking can be done manually on a directory of modest size.
What does require developer support, even with the most permissive CMS, is the segmented sitemap by entity type, the custom referrer-tracking pipeline for AI sources, and any non-trivial scripted intervention (the embeddings-driven internal linking, for instance). Solo operators have to either accept these gaps, partner with a freelance developer for tightly scoped engagements (a competent freelancer can usually deliver sitemap segmentation and referrer tracking in two or three days of focused work, costing perhaps $2,000), or use vendor-supplied tooling that handles these elements at a layer above the CMS.
The principle here, drawn from the comprehensive review of optimisation methods in Artificial Intelligence Review (Springer, 2024), is that simpler approaches with fewer control parameters often produce more reliable results than sophisticated methods that require ongoing tuning. A solo operator executing the seven or eight tactics that fit their constraints, well, will often outperform a better-resourced competitor executing all twelve poorly. The full programme is not a checklist where partial completion produces partial results; it is an ordered set where the right subset, sequenced correctly, captures most of the value.
Regulated Industries and Compliance Limits
Directories in regulated industries — financial services, healthcare, legal services, certain regulated consumer categories — face constraints that the SaaS engagement did not encounter. The tactics still apply but require modification, and one tactic in particular (the answer-ready snippets) has to be handled with substantial care.
The compliance constraint that bites hardest is the prohibition on giving advice, even implicitly, in jurisdictions where the directory is not licensed to do so. A snippet that answers “what is the best mortgage for a first-time buyer in California?” with a substantive recommendation is, in some regulatory readings, a regulated act. Directories operating in these spaces have to reframe their snippets as factual descriptions of features rather than recommendations — “[Lender X] offers first-time buyer mortgages with the following terms…” rather than “[Lender X] is good for first-time buyers because…”. The semantic difference is small; the legal difference is not.
The same caution applies to aggregateRating in regulated categories. Healthcare directories in particular have to be careful about how patient ratings are aggregated and displayed — there are jurisdictions where unmoderated ratings can create liability, and the compliance overhead of moderating them at scale can be substantial. Some directories in these spaces choose to omit ratings entirely, which weakens the schema’s effectiveness but eliminates the risk.
Citation tracking also looks different in regulated spaces. AI systems are increasingly cautious about providing specific advice in regulated categories, often refusing to answer or hedging heavily. A healthcare directory may find that its citation rate in generated answers is lower than a SaaS directory’s not because its content is weaker but because the systems themselves are reluctant to engage with the queries. This is not a problem the directory can solve; it is a structural feature of the environment, and tactical investment should be calibrated accordingly.
The redeeming factor in regulated industries is that the bar for becoming the canonical structured-data source for the vertical is often lower, because fewer competitors are willing to do the compliance work. A regulated-industry directory that invests in clean, conservative, accurately-marked-up listings can establish a position that is difficult for newer entrants to dislodge, even if the short-term traffic gains are smaller than in less regulated spaces. The Deloitte Insights framing of cost optimisation as growth-oriented applies particularly well here — the investment in compliance-aware structured data is best understood as building durable infrastructure rather than chasing immediate visibility.
Several questions have surfaced through this work that the available evidence cannot yet resolve, and that future research — whether by academic groups, industry analysts, or practitioners with longer measurement windows — would do well to address. The first is whether AI-mediated citation behaviour is durable or transient: do the gains observed in this engagement reflect a lasting structural advantage, or are they an early-mover effect that will compress as more directories adopt similar tactics? Three to five years of longitudinal data across multiple verticals would clarify this, and to date no such dataset exists in the public domain. The second is how AI retrieval systems will evolve their citation logic as their training corpora become saturated with content explicitly engineered for citation; there is a plausible scenario in which the systems develop counter-strategies that penalise the very tactics that currently benefit, and the field would benefit from rigorous adversarial analysis of this dynamic. The third concerns the threshold at which the unbundling of directories by AI mediation becomes existential rather than merely uncomfortable: at what share of mediated discovery does the directory business model break, and which sub-categories of directories are most exposed? These are not questions that any single engagement, however well-instrumented, can answer — but they are the questions that will shape the next decade of this category, and the practitioners and researchers willing to investigate them seriously will do the field a substantial service.
