Wikipedia held the top SEO Visibility score across the U.S. Google index in 2018, according to Statista’s Searchmetrics-derived ranking, a directory-like structure outperforming every commercial brand on the web. That single data point, now several years old, has quietly become one of the most relevant signals in the generative AI era. Here is why. The same structural traits that lifted Wikipedia in classical search, dense entity relationships, consistent factual scaffolding, and machine-readable formatting, are exactly what large language models prefer when they decide which sources to cite. The directory pattern was already winning before anyone said the phrase “AI search visibility.”
What follows is a walkthrough of a real engagement (composited and anonymised across three clients to protect commercial confidentiality) where directory work showed measurable effects on AI citations. The reasoning at each fork matters more than the specific brand involved, so I will stay on the decisions, not the name-dropping.
The client problem that started this
The client was a mid-market B2B SaaS company in the workforce analytics space, roughly $14M ARR, 62 employees, operating across the U.S. and the U.K. They came in with a complaint that has become depressingly common since mid-2024. Their organic traffic from Google was holding steady (down only 4% year on year, which counts as a win in the post-AI-Overviews era), but their inbound demo requests attributed to “ChatGPT mentioned you” or “Perplexity recommended you” were essentially zero. Worse, when their CEO ran a series of vanity prompts during a board meeting, “best workforce analytics platforms for mid-market HR teams,” the brand was absent from every AI engine tested. Two direct competitors were named consistently. One of those competitors had roughly half the brand’s organic traffic and a smaller content team.
The marketing director’s framing was honest: “We have no idea what these systems read, and we have no idea why they don’t read us.” That framing matters because we negotiated the engagement scope around an investigation, not a guarantee. Anyone selling guaranteed AI citation lift in 2025 is selling something that does not exist yet. Going in, the hypothesis was that the brand had a citation supply problem, too few high-authority third-party pages mentioning the company in entity-rich contexts, and that targeted directory work, alongside structured data cleanup, could begin to close that gap within a quarter.
Budget was modest by enterprise standards: roughly $18,000 across ninety days, covering audit, listing creation, copy work, and measurement. No paid placements, no pay-to-play directories. The client also imposed a constraint that turned out to be useful: anything published on their behalf had to be factually defensible to their compliance team, because workforce analytics touches HR data and the legal exposure of an inaccurate third-party listing is real. That constraint forced discipline. Half the poor directory hygiene I have seen across 200-plus audits would not survive a compliance review.
Setting the baseline for AI visibility
Auditing current citation sources
The audit began with a citation inventory rather than a backlink inventory, because those are different things. A backlink is a hyperlink; a citation is a third-party mention of the brand, with or without a link, in a context that names the entity clearly. LLMs ingest both, but the unlinked mention is often more useful to them than SEOs realise. The training corpora that produced GPT-4-class models do not require a hyperlink to associate “Acme Workforce Analytics” with the entity attributes “HR software,” “mid-market,” and “Salt Lake City.”
The inventory used a combination of Ahrefs (for the linked footprint), Brand24 and Mention (for unlinked references in news, blogs, and forums), and a manual sweep across 28 well-known business directories. The audit catalogued 217 distinct third-party pages mentioning the brand. Of those, only 41 contained what I would classify as an entity-rich description, meaning a passage of at least 40 words connecting the brand name to its product category, ideal customer profile, and at least one differentiator. The rest were name-only citations, broken listings, or pages where the brand appeared as a single line item in a long aggregated list.
Forty-one entity-rich citations is, in my experience, below threshold. The brands that perform well in AI engines tend to sit somewhere between 120 and 400 such citations across the open web before LLMs start surfacing them as default recommendations. That number is empirical rather than published. It comes from running similar audits on roughly thirty brands over the past eighteen months, but the directional signal is consistent.
Mapping competitor directory footprints
Two competitors were named repeatedly in AI responses. The audit reverse-engineered their citation footprints using the same toolset. The pattern was immediate: the named competitors had presences on G2, Capterra, TrustRadius, GetApp, Software Advice, SaaSworthy, and roughly fifteen vertical HR-tech listings (SHRM partner pages, HR Executive resource hubs, HR Tech Alliance member listings, and similar). Each of those profiles ran 200 to 600 words of branded copy, structured Q&A sections, and customer review bodies that themselves used entity language.
The client had presences on three of those properties. Two were stale (last updated in 2022). One was claimed but unfilled, a shell page with the company name and nothing else. The competitors had, in effect, built a redundant network of authoritative pages that any reasonable web crawler, including those operated by OpenAI, Anthropic, Perplexity, and Google’s Gemini infrastructure, would encounter while building or refreshing their indices.
Identifying the AI citation gap
ChatGPT brand mention tests
I ran a standardised prompt set against ChatGPT (GPT-4o and GPT-4-turbo at the time of the audit), repeated across three sessions, with web browsing both enabled and disabled. Twenty-five prompts covered category queries (“workforce analytics platforms for companies with 500-2000 employees”), problem queries (“how do I measure employee productivity without surveillance concerns”), and direct comparison queries (“alternatives to [Competitor A]”). The brand appeared zero times in browsing-disabled mode and twice in browsing-enabled mode, both times as a low-priority third-tier mention buried in a long list. Competitor A appeared in 21 of 25 prompts. Competitor B appeared in 17.
Perplexity source tracing
Perplexity is more useful for diagnosis than ChatGPT because it shows the citations it draws from. Running the same prompt set, Perplexity cited G2 (eight times), Capterra (six times), a TechCrunch article from 2023 (three times), and a clutch of HR trade publications. The client’s own domain was cited once across 25 prompts, for a single product page that happened to rank well organically. The competitors’ G2 profiles were cited repeatedly. The mechanism was visible in the source list: Perplexity was treating G2 entries as canonical descriptions of each product.
Google AI Overview patterns
Google’s AI Overviews are noisier and less reproducible (the same query returns different overviews across sessions, geographies, and account states), but a sample of fifty queries surfaced a consistent pattern. AI Overviews preferred sources that combined editorial authority with structured product data. Forbes Advisor’s vendor roundups, G2’s category leaders pages, and SHRM’s resource hubs appeared far more often than vendor-owned content. The brand’s own marketing pages were essentially invisible in this surface.
The diagnosis at the end of baseline week was straightforward: the brand was undercited on the third-party properties that AI engines treat as authoritative summary sources for the category. The remedy was directory and listing work, but only the kind of directory work that survives modern crawler heuristics.
Selecting directories that AI engines trust
Why domain authority still matters
Domain authority (or its various proxies: Ahrefs DR, Moz DA, Majestic Trust Flow) is not a Google ranking factor and never has been, but it remains a useful filter when selecting directories for an AI visibility programme. The reason is indirect: high-authority domains tend to be crawled more frequently, retained longer in training corpora, and referenced more often by other authoritative sites. LLMs trained on Common Crawl snapshots inherit these biases. A listing on a DR 85 directory has a meaningfully higher probability of ending up in a model’s training data than a listing on a DR 22 directory, even if neither directly contributes a Google ranking signal.
The Statista visibility data showing Wikipedia’s dominance is instructive here, even though it is from 2018. The structural reason Wikipedia performs well in classical search is the same structural reason it performs well in LLM citation: dense linking, factual consistency, and schema-friendly formatting. Directories that mimic those properties, such as strong internal linking, consistent entity attributes, and machine-readable markup, inherit some of the same benefit.
Niche versus general directories
The instinct of inexperienced practitioners is to chase general directories first because they sound bigger. That instinct is wrong for AI visibility purposes. A vertical directory with 4,000 entries in a single category gives a brand an entity context that a general directory with 4 million entries across every category cannot match. When an LLM is asked about workforce analytics, the model is more likely to surface sources where the surrounding text is densely about workforce analytics, because that co-occurrence is what trained the embedding in the first place.
That said, general directories still earn a place in the stack, but as anchor citations that establish baseline NAP (Name, Address, Phone) consistency, not as the primary AI visibility play. The split in this engagement landed at roughly 25% general directories, 60% vertical or niche, and 15% long-tail or geographic. That ratio has held up across other engagements too.
Checking crawl and indexation signals
Before any directory was added to the shortlist, I ran three checks on it. First, was the directory’s content indexed in Google? A surprising number of older directories now sit behind soft paywalls, JavaScript walls, or robots.txt blocks that exclude them from major crawlers. Second, did the directory return clean HTML to a non-JavaScript request? LLM training pipelines vary, but many still rely on text extraction from server-rendered HTML. A directory whose listings only render after client-side JavaScript execution is, for current model purposes, partially invisible. Third, was the directory accessible to the named AI crawlers, GPTBot, ClaudeBot, PerplexityBot, and Google-Extended? A growing number of properties block one or more of these, sometimes deliberately, sometimes by accident through overzealous WAF rules.
Roughly 40% of the candidate directories failed at least one of these three checks. That filter alone saved the engagement from wasting time on properties that, however reputable, would not have moved the AI visibility needle.
The three-tier directory stack
Tier one anchor listings
Tier one consisted of seven properties: G2, Capterra, TrustRadius, GetApp, Software Advice, Crunchbase, and LinkedIn’s company directory. These are non-negotiable for any B2B SaaS programme. They are also the properties that AI engines cite most frequently in this category. The work here was not creation but rebuild: full profile completion, fresh customer reviews solicited through a structured outreach campaign, and category placement reviews. Tier one consumed roughly 45% of the engagement’s effort budget.
Tier two vertical directories
Tier two pulled in HR-specific properties: SHRM resource pages, HR Executive’s vendor directory, HR Tech Alliance member pages, People Managing People’s vendor lists, Select Software Reviews, and roughly twelve trade publication resource hubs. Several of these required editorial review before listing approval, which lengthened the timeline but raised the quality of the resulting page. Tier two consumed about 35% of effort.
Tier three long-tail plays
Tier three was the experimental stack: emerging curated lists, regional business directories in the client’s two operational markets, AI tool directories that had recently added “HR” or “workforce” categories, and several general business directories with strong domain trust signals and clean indexation. This is where curated general-purpose listings come in. As documented in a recent piece on entity-first listing structures, the directories that perform best for AI surfaces tend to enforce editorial review and structured field consistency rather than allowing free-form vendor-written copy. Tier three consumed the remaining 20% of effort and was treated as a hedge: low individual probability of impact, useful in aggregate.
Directories we decided to skip
Low-quality link farms
Anything that looked like a 2010-era SEO link farm (auto-approved listings, unedited submission queues, scraped content from other directories) was excluded. These properties don’t help with traditional SEO anymore, and they are systematically deprioritised or filtered out of LLM training data. Submitting to them is, at best, a waste of time. At worst, it associates the brand with low-quality neighbourhoods in ways that are difficult to undo.
Paywalled or closed sources
Several otherwise reputable directories, including a couple of Gartner-owned properties, sit behind paywalls or registration walls that prevent open crawling. These can still be valuable for human buyers in late-stage research, but their contribution to AI visibility is essentially zero because the content never enters the open web’s training corpora. They were excluded from this programme on cost-effectiveness grounds, with a note to revisit if the client’s budget expanded.
Directories blocking AI crawlers
A small but growing category. Some directories, particularly those that monetise their data through API licensing, block GPTBot, ClaudeBot, and similar agents in robots.txt. Listing on these directories may still help with traditional SEO and human discovery, but for AI visibility specifically, they are dead ends. The client’s compliance team appreciated this filter, because it also reduced exposure to properties whose terms of service were ambiguous about how listings might be redistributed.
Final shortlist of forty directories
The final shortlist contained 40 directories: 7 tier-one anchors, 18 tier-two verticals, and 15 tier-three long-tail entries. The original candidate list had been 94. The cull was aggressive, and that aggression is, on balance, the most underrated practitioner skill in this work. The temptation to add “one more” directory is constant, and almost always wrong. Forty was already at the upper edge of what could be maintained quarterly within the client’s operational capacity.
Crafting listings for LLM consumption
Writing entity-rich descriptions
Listings written for human readers and listings written for LLM consumption are not the same artefact, although the difference is subtler than most copywriters assume. The human-optimised listing leans on benefit language, social proof, and rhetorical hooks. The LLM-optimised listing leans on entity density: what the product is, who it serves, what category it sits in, what it integrates with, what differentiates it from named alternatives, and what verifiable facts about it (founding year, headquarters, customer count, certifications) ground its identity in the model’s representation.
The compromise this engagement settled on was a layered structure. The first 80 words read naturally for a human evaluator; the following 200 to 400 words systematically covered the entity attributes that LLMs use to disambiguate brands. Each listing referenced two or three named competitors in a comparison context, not as bashing, but as factual positioning (“designed for mid-market HR teams that find [Competitor A] too enterprise-heavy and [Competitor B] too lightweight on compliance reporting”). Co-occurrence with named competitors is, in practice, one of the strongest signals an LLM uses to place a brand in a category.
Structured data and schema choices
Where the directory permitted custom HTML or schema injection (relatively few do), the listings used Organization, SoftwareApplication, and Product schemas with consistent property values across every directory. Where the directory enforced its own schema (as G2 and Capterra do), the focus shifted to making sure the structured fields were complete and consistent: pricing model, deployment options, target company size, integrations, supported languages. These structured fields are the cleanest signal a directory can pass to a downstream LLM crawler, and inconsistencies across directories actively confuse model representations.
One specific tactic that paid off disproportionately: filling out the “alternatives to” and “competitors” fields on every property that exposed them. Those fields create explicit graph edges between the brand and competing brands, and those edges are precisely what models traverse when answering “alternatives to” prompts.
Consistent NAP and brand facts
Name, Address, Phone consistency is local SEO 101, but its role in AI visibility is underappreciated. When a model sees the same brand name attached to three different addresses across three directories, the most charitable interpretation it can form is uncertainty. The less charitable interpretation is that these are three different entities. Either way, the model’s ability to confidently surface the brand in response to category queries degrades.
The cleanup pass standardised brand name formatting (one canonical form, no abbreviations or trailing “Inc.”), single registered address, single primary phone, single founding year, single employee count band, single funding round disclosure. The client’s compliance team verified each fact before publication. This step took eight working days and was, by mid-engagement, recognised as the highest-leverage individual workstream in the entire programme.
Rolling out submissions over six weeks
The rollout was paced rather than blitzed, for two reasons. First, several tier-two and tier-three directories require editorial review, and overlapping fifteen review processes simultaneously creates a coordination nightmare when revisions come back. Second, a sudden flood of new third-party citations referencing a brand can, in some edge cases, trigger spam-detection heuristics on aggregator sites. A six-week rollout creates a more natural citation velocity profile.
Week one focused on the seven tier-one anchors. The G2, Capterra, TrustRadius, and GetApp profiles were rebuilt from scratch, including new screenshots, video walkthroughs, updated feature lists, and category placement reviews. Crunchbase and LinkedIn were updated with the canonical NAP and brand facts. Software Advice was negotiated separately because it shares infrastructure with Capterra and Gartner Digital Markets, so coordinating updates avoided duplicated effort.
Weeks two and three covered the eighteen tier-two vertical directories. SHRM’s resource page required a partner agreement that took eleven days to finalise. HR Executive’s vendor directory required editorial review of the listing copy, which came back with two rounds of revisions before approval. Several smaller HR trade publications had simpler workflows but inconsistent submission interfaces. A few still operate via email submission rather than web forms, which slows things down considerably. The team learned to start tier-two submissions before tier-one was fully complete, simply because the editorial calendars on these properties are unpredictable.
Weeks four and five covered the fifteen tier-three long-tail entries. These were faster to execute individually but heavier in aggregate volume. Quality control here mattered more than speed, because tier-three properties are exactly where rushed work shows up as inconsistent NAP or copy-pasted descriptions that read identically across multiple directories. Identical copy across directories is a signal LLMs use to discount sources, since repetition of identical text suggests low editorial value. The team rewrote each tier-three description from scratch, reusing the same factual scaffolding but varying the prose meaningfully.
Week six was reserved for review and remediation. By that point, most submissions were live, several had returned editorial questions, and a handful had silently failed without notification. (One property accepted the submission, generated a confirmation email, and never actually published the listing. A manual check four weeks later caught the omission.) The remediation week is non-negotiable. Practitioners who skip it routinely discover three months later that 15 to 20% of their submissions never went live.
Measuring AI visibility lift
Tracking citation frequency
Manual prompt testing
The same 25-prompt set used in the baseline was re-run weekly against ChatGPT, Perplexity, Claude, and Gemini. Manual prompt testing remains the most reliable measurement method for AI visibility, despite being labour-intensive. It captures nuance, whether the brand is mentioned as a primary recommendation, a third-tier alternative, or a passing reference, that automated tools struggle to extract reliably. Each test session was logged with screenshots, the model version (where exposed), and the cited sources where the engine surfaced them.
Automated monitoring tools
Several tools have entered the market in the past eighteen months claiming to track AI citation frequency at scale: Profound, Otterly, AthenaHQ, and a handful of others. The engagement used Profound for trend tracking, but treated its outputs as directional rather than authoritative. Automated tools rely on API-level access where it exists and on prompt automation where it doesn’t, and the prompts they run are not always representative of how real buyers query these systems. They are useful for spotting major movements between weeks but should not be the sole measurement instrument.
Numbers after ninety days
ChatGPT mention increase
Across the 25-prompt set, the brand moved from 0 mentions (browsing-disabled) and 2 mentions (browsing-enabled) at baseline to 6 and 14 mentions respectively at ninety days. The browsing-disabled increase is the more notable signal, because it reflects what has entered the model’s parametric knowledge or, where applicable, what the model retrieves from cached training-period sources. The browsing-enabled increase is partly a reflection of better retrieval surfaces: the new G2 and Capterra content is now what the model finds when it searches.
Perplexity source inclusions
Perplexity’s diagnostic value showed up clearly. The brand’s domain was cited 7 times across the 25-prompt set at ninety days, up from 1 at baseline. More importantly, the third-party properties cited alongside the brand had shifted: G2 (now 14 citations, up from 8), Capterra (10, up from 6), three HR trade publications that had not appeared at baseline, and two SHRM resource pages that had become standard sources for category queries. The model was, in effect, treating the rebuilt directory ecosystem as the canonical answer surface for the category, and the brand was now visible inside that surface.
Referral traffic changes
Referral traffic from AI engines is notoriously hard to attribute, because the referrer headers from ChatGPT, Perplexity, and Gemini are inconsistent and sometimes stripped entirely. The cleanest signal came from a custom UTM strategy that tagged links published in the directory listings themselves, allowing direct attribution of clicks from those properties. Referral traffic from the directory stack rose from 412 sessions per month at baseline to 1,847 sessions per month at ninety days. Of those, an estimated 22% (based on landing page behaviour analysis and time-of-day patterns) appeared to originate from AI-engine workflows, users clicking through after a model surfaced the brand.
Demo requests attributed to “found you through ChatGPT/Perplexity” via the post-signup survey rose from 1 in the ninety days before the engagement to 23 in the ninety days during. Twenty-three demo requests is not a large number in absolute terms, but for a $14M ARR business with a $48,000 average contract value, it represented a meaningful pipeline contribution against an $18,000 engagement cost.
What surprised us most
Two things. First, the speed at which Perplexity reflected the new citation surface. Meaningful changes appeared within three weeks of the tier-one rebuilds, faster than expected. ChatGPT lagged by several weeks because its retrieval-augmented behaviour weights cached representations more heavily. Second, the disproportionate impact of the “alternatives to” fields on competitor comparison prompts. A single G2 listing change, completing the alternatives field with three named competitors, appeared to be responsible for a measurable share of the comparison-prompt mentions. The graph edges those fields create matter more than any single prose paragraph.
Transferable principles from this engagement
Treat directories as training data
The mental model shift that separates practitioners who get results from those who don’t is treating directory listings as training data, not as marketing pages. Every published word will, eventually, end up in some model’s representation of the brand. That framing changes the editorial bar. It rewards factual precision over rhetorical flourish, structured fields over free-form copy, consistency over creative variation. The Forrester observability framing is useful here: Forrester’s analysis of observability argues that visibility in complex systems requires combining logs, metrics, and traces, separate signal types that together form a coherent picture. AI visibility behaves the same way. No single directory listing creates visibility; the aggregate signal across structured listings, schema markup, customer reviews, and editorial mentions creates the entity that models can reliably surface.
Prioritize entity consistency over volume
Forty well-maintained, factually consistent listings outperform 200 inconsistent ones. The audit work on the brand’s existing 217 third-party citations revealed that several of the older listings carried conflicting facts: a 2019 funding round amount that had since been superseded, a previous office address, an outdated employee count, a former product name. Each of those inconsistencies created uncertainty in any model that ingested both the old and new data. The cleanup pass on stale listings was, in retrospect, as valuable as the new submissions. Findings from Deloitte’s insights indicate that only 21% of enterprises report mature governance for agentic AI deployments, and the same governance gap shows up in entity consistency. Most brands have no single source of truth for their own facts across the open web, and that absence is what models inherit.
Adjusting the playbook for different constraints
Running this on a smaller budget
A $5,000 version of this engagement would compress to the seven tier-one anchors plus three or four high-impact tier-two verticals, and would skip the tier-three long-tail entirely. The compression sacrifices roughly 30% of the achievable lift, in my estimate, but preserves the highest-leverage workstreams: NAP cleanup, G2 and Capterra rebuilds, and entity-rich descriptions on the most-cited vertical directories. The work that should never be cut, even at $2,000, is the entity consistency pass. That is foundational, and skipping it makes everything downstream weaker.
According to a study available here on lean directory programmes, the marginal return on directories beyond the top fifteen is often a fraction of the return on the first ten, a pattern that supports concentrating limited budgets at the top of the stack rather than spreading them thinly.
Adapting for regulated industries
Healthcare, financial services, and legal services impose constraints that change the playbook materially. Several directories are off-limits for regulated brands because their editorial controls are insufficient to prevent misleading claims (FDA, FCA, and state bar regulators all care about this). The substitute strategy leans more heavily on professional association directories, regulator-recognised registries, and editorial publications that maintain compliance review. Effort shifts from quantity to depth: fewer directories, more thorough listings, more emphasis on the verifiable facts that regulators care about (license numbers, registered addresses, named officers). The timeline also extends, often by 30 to 60%, because every listing requires legal sign-off.
Research published by Harvard Business Review on AI adoption notes that empathetic, governance-aware deployment reduces resistance and accelerates impact, a principle that translates directly to regulated-industry directory work, where the compliance team’s involvement should be treated as a structural feature, not a friction point.
Compressing the timeline to thirty days
A thirty-day version is possible but requires sacrificing tier-three entirely and running tier-one and tier-two in parallel rather than sequentially. The risk profile rises sharply: editorial revisions on tier-two properties may not return within thirty days, and chasing them creates submission errors that cost more time downstream. The thirty-day version also forfeits the natural citation velocity profile that paced rollouts produce. In one prior engagement run on a compressed timeline, two tier-two properties rejected the listing on first submission for what amounted to formatting errors that would have been caught in a more deliberate process. Those rejections then required a thirty-day cooling-off period before resubmission. The total elapsed time was longer than a properly paced rollout would have been.
Mistakes to avoid on your first run
Chasing every directory available
The single most common mistake is treating directory submission as a volume game. It isn’t. The marginal value of the fortieth directory is small; the marginal value of the eightieth is essentially zero, and the maintenance burden it creates is positive. Practitioners who chase every directory available end up with a sprawling footprint they cannot maintain, which decays into exactly the inconsistent-facts problem the cleanup pass was supposed to solve. Forrester’s CX research on the “line of visibility” frames this well: the failure mode is unintentional exposure of ecosystem elements that should have remained hidden, alongside hiding elements that should have been visible. Directory sprawl produces both failures simultaneously.
Ignoring listing maintenance
A listing published in March and never touched again decays. Customer reviews accumulate that may be outdated. Product features change. Pricing tiers evolve. Competitors update their own listings, shifting the comparison context. A quarterly maintenance cycle, even a lightweight one, keeps the listing fresh in ways that LLM crawlers reward. The engagement built a quarterly review checklist that took roughly four hours per quarter to execute across the forty-listing footprint. That is not a heavy ongoing burden, but it is a non-zero one, and brands that pretend otherwise discover their carefully built footprint has decayed within eighteen months.
Skipping pre-launch entity cleanup
Submitting new listings before cleaning up the existing ones is the canonical rookie error. It actively makes the situation worse, because new listings with the corrected facts now contradict twenty older listings with the wrong facts. The model sees more conflict, not less. The pre-launch cleanup is unglamorous work: chasing down stale listings, updating funding amounts, correcting addresses, requesting removals where appropriate. Skipping it undermines everything that follows.
Misreading early citation noise
The first three weeks of any AI visibility programme produce noise. Models update on irregular cadences. Retrieval augmentation introduces session-to-session variability that can look like signal but isn’t. Practitioners who declare victory or defeat at week two are reading noise. The earliest defensible read on programme impact is around week six, and the most reliable read is at week twelve. McKinsey’s 2024 Technology Trends Outlook notes that generative AI deployment patterns frequently produce non-linear adoption curves, and the same non-linearity applies to AI search visibility. Linear extrapolation from week-two data is a mistake the field continues to repeat.
Three open questions surface from this engagement that the field has not yet resolved, and that future research should pursue. First, what is the half-life of an AI citation gain: how quickly does directory-driven visibility decay when maintenance lapses, and does that decay rate vary across model families and retrieval architectures? Second, how do model-specific retrieval behaviours (parametric memory in ChatGPT versus real-time retrieval in Perplexity versus the hybrid approach in Google’s AI Overviews) reward different directory strategies, and is there a unified strategy that performs adequately across all three, or do practitioners face genuinely irreducible trade-offs? Third, as more directories block AI crawlers in robots.txt or pursue licensing arrangements with model providers, will the open web’s directory ecosystem bifurcate into a public tier (visible to all models) and a licensed tier (visible only to paying providers), and what would that split mean for brands that cannot afford listings in both? Each of these is answerable empirically, with enough longitudinal data and enough collaborating practitioners willing to share results. The field is still young enough that the answers are not yet settled, and the practitioners who treat this as an ongoing investigation rather than a solved playbook will, in my experience, continue to outperform those who have already declared the question closed.
