- An AI visibility strategy is a five-stage dependency chain, not a checklist: access & indexation → entity consistency → content extractability → earned off-site presence → ongoing monitoring. Each stage gates the next.
- A one-time "AI visibility score" isn't a baseline. SparkToro found under a 1-in-100 chance that two runs of the identical prompt return the identical brand list.
- llms.txt (97% of files get zero AI-crawler requests, per Ahrefs' 137,000-domain study) and schema markup (no measurable AI-citation lift, per Ahrefs' 1,885-page study) are the two most oversold tactics in AI search optimization.
- Earned media (not owned content) drives 82–89% of all AI citations across three independent Muck Rack studies since July 2025.
- Run these stages out of order and you burn budget: off-site PR pointed at a crawler-blocked site, or a monitoring dashboard measuring a brand nobody can find.
Most of what's sold as "AI search strategy" right now is a repackaged audit with a scarier name. Someone runs your brand through ChatGPT a few times, screenshots whether you showed up, and hands you a PDF with a visibility score on the cover. You pay for it, feel briefly reassured or briefly alarmed, and nothing structural changes.
That number was never a foundation. It's a snapshot of noise, and the research now proves it, precisely.
SparkToro's Rand Fishkin and Gumshoe.ai's Patrick O'Donnell ran 2,961 prompts across ChatGPT, Claude, and Google's AI surfaces in November and December 2025, using 600 volunteers repeating the same 12 prompts 60–100 times per platform. The odds of two runs returning the identical brand list: under 1 in 100. The odds of the identical list in the identical order: under 1 in 1,000. A single check of "are we in ChatGPT" tells you almost nothing about whether you're actually in the consideration set. It tells you what one dice roll looked like.
That single fact should end the one-time-audit business model. It hasn't, because inconsistency is easier to sell fear around than to fix. So here's the alternative: the five stages for winning in the AI search era, run in a fixed order, each one gating the next. This isn't a framework I'm proposing in the abstract. It's the sequence I run as a fractional AEO/GEO/SEO strategist, built on the AI-first content system I ran in-house at Atlan, and stress-tested against the primary research published through mid-2026.
Why did the AI search era need a different playbook in the first place?
Three shifts, all measurable, all happening at once:
The platform you're optimizing for keeps changing. Similarweb's July 2026 data shows ChatGPT's share of worldwide generative-AI web traffic falling from roughly 76% in June 2025 to about 53% by May 2026, not because usage dropped, but because Gemini crossed a quarter of category traffic and Claude became the fastest-growing platform in the set. Building a strategy around one model's citation habits is building on sand that's actively shifting under you.
Clicks are disappearing at the query level, not just the SERP level. By multiple independent 2026 trackers, AI Overviews now appear in somewhere between 18% and 26% of Google searches depending on the study and the month, and click-through on a query with an overview present roughly halves versus one without. That's not a hypothetical future. It's the current baseline you're already operating inside.
The answer itself is unstable. The SparkToro finding above isn't an edge case: Ahrefs separately found Google's AI Mode and AI Overviews cite different sources 87% of the time for the identical query, on the identical platform. Two different studies, two different mechanisms of instability, same conclusion: point-in-time visibility checks are structurally unreliable as a management metric.
Put those three together and the old playbook (rank a page, track a position, report the number) doesn't map onto the terrain anymore. What replaces it isn't more tracking. It's sequencing: fixing what's binary before you spend a rupee on what's probabilistic.
What is the sequence?
Five stages. Each one has to clear before the next one is worth starting. Running them out of order is the single most common way I've seen budget get burned in this category: paying for off-site earned-media work while the site is invisible to crawlers, or buying a monitoring subscription before there's anything worth monitoring.
| Stage | Gate type | What breaks if skipped |
|---|---|---|
| 1. Access & indexation | Binary gate | Nothing downstream produces citations: a crawler that can't reach you can't cite you |
| 2. Entity consistency | Resolution layer | Signals (reviews, press, stats) can't reliably attach to the right brand |
| 3. Content extractability | Highest-evidence lever | Content exists but isn't in a shape a model can quote |
| 4. Off-site & earned presence | Highest-ceiling, slowest | Growth stalls: 82–89% of citations come from here, per Muck Rack |
| 5. Ongoing monitoring | Measurement only | You're tracking noise instead of a trend, per SparkToro's consistency data |
Stage 1: is access and indexation in place?
What it is: Confirming AI crawlers can physically reach your site, and that you're indexed in the index each engine actually retrieves from. ChatGPT and Copilot draw heavily on Bing's index. Google's AI Overviews and AI Mode draw on Google's own index. If either gate is closed, nothing downstream matters. You don't get partial credit for great content a crawler never fetched.
Why it's first, non-negotiably: This is a gate, not a lever. A brand with perfect entity hygiene and a dozen glowing G2 reviews still produces zero AI citations if a firewall rule or a robots directive is silently blocking the crawler. I've hit this exact failure on my own domain: a Cloudflare managed-rules bot block that was silently overriding a permissive robots.txt. The site looked fine to a human. It was invisible to the systems that mattered, and nothing in the analytics said so directly; you had to go read server logs to find it.
What the evidence says to actually check here, versus what gets sold instead: This is also where most of the industry's manufactured busywork lives. Two 2026 studies are worth knowing cold before anyone bills you for "AI crawler optimization":
- llms.txt. (The full standalone answer to whether it works at all is in does llms.txt actually get read by anyone?; Google has since said in writing that it ignores the file entirely.) Ahrefs analyzed 137,210 domains with server-log access in June 2026. 97% of published llms.txt files received zero requests in the entire month of May. Where AI retrieval bots (the ones that actually generate citations, as opposed to training crawlers or coding agents) did request anything, they accounted for 1.1% of total requests to the file. The largest single category of requester was SEO audit tools checking whether the file existed, not AI systems reading it. SE Ranking's separate 300,000-domain study found no statistically significant correlation between having an llms.txt file and AI citation frequency at all: removing the variable from their predictive model improved its accuracy. None of OpenAI, Google, Anthropic, Meta, Perplexity, or Mistral has stated that production retrieval systems act on the file.
- Schema markup. Ahrefs ran a matched difference-in-differences study across 1,885 pages that added JSON-LD schema between August 2025 and March 2026, against 4,000 control pages with similar pre-treatment citation levels. Result: statistically indistinguishable from zero on ChatGPT and Google AI Mode, and a small but statistically significant decline on Google AI Overviews. A separate live-fetch test published in Search Engine Journal in June 2026 found that when an answer engine pulls a page in real time, it frequently ignores the JSON-LD entirely and reads the visible HTML text. The one nuance worth keeping: a February 2026 cross-platform study found schema populated with genuinely extractable facts (concrete pricing, aggregate ratings, specs) correlated with meaningfully higher citation rates than generic Organization or BreadcrumbList schema. The lift isn't in having schema. It's in the underlying data being concrete and quotable, which schema can expose but doesn't create.
Stage 1, done properly, is server-log analysis (which crawlers are hitting you, at what rate, blocked or served), a Bing Webmaster Tools and Google Search Console indexation check, and a robots.txt / WAF / CDN audit specifically for the AI-bot user agents (GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended), not a file nobody reads. For the full walkthrough, see how to check AI crawler access, the deeper user-agent and IP verification methodology, and how to turn a log check into a repeatable report.
Stage 2: is your entity consistent?
What it is: Making your brand resolve as one clean, disambiguated entity across the sources a model cross-references to decide who you are: Wikidata, consistent legal and brand naming across your own domain, G2 and Capterra profiles that agree with each other, LinkedIn and Crunchbase data that doesn't contradict your own site.
Why it comes second: A model doesn't cite a name. It cites an entity it has some confidence it has correctly resolved: the right company, the right category, the right facts attached. If your brand shows up under three slightly different names across your own footprint (a legal entity name on your privacy page, a product name on your homepage, an old name still live on a partner directory), you're asking the model to do disambiguation work it has no strong incentive to finish. Get this wrong and every downstream signal (a review, a press mention, a stat) has a real chance of attaching to the wrong node, or not attaching at all.
What actually moves this: A Wikidata entry if you don't have one and qualify for one. One canonical name used identically everywhere you control. Closing the gap between what your own site says about you and what third-party profiles (G2, Capterra, Crunchbase, LinkedIn) say. This is unglamorous, mechanical work, and it's the reason Stage 2 sits before Stage 3, not after: extractable content about a poorly-disambiguated entity is still poorly-disambiguated content. For what this actually means and why it matters, see what is entity consistency; for the mechanical checklist, entity audit in under an hour; and for the two highest-leverage profiles specifically, Wikidata for brands and G2/Capterra profile hygiene.
Stage 3: is your content extractable?
What it is: Restructuring pages so a model can lift a direct, self-contained answer from them: clear direct answers near the top of a section, claims that stand alone without requiring the surrounding paragraph for context, original data a model can't get from three other sites that all cite each other.
Why this is the one on-site lever worth real investment: Of everything covered here, this is the stage with the strongest research base specifically because it's the one variable multiple independent studies actually isolated and tested, rather than assumed. The often-cited Princeton/Georgia Tech GEO paper (arXiv 2311.09735), frequently misused to justify schema markup, which it never tested, found that adding citations, statistics, and direct quotations to content measurably improved a source's visibility in generative answers. Critically, the paper is explicit that the effect is domain-dependent: tactics that outperform in factual categories (science, legal, historical) underperform in others (opinion, politics). There is no universal checklist here; the paper's own data contradicts a one-size-fits-all approach.
What "extractable" means in practice, concretely: A direct answer in the first 1–2 sentences of a section, not buried after three paragraphs of throat-clearing. Original numbers (your own benchmark, your own survey, your own before/after) rather than a paraphrase of someone else's stat, because a model has no reason to cite the tenth restatement of the same figure over the primary source. Self-contained sections that answer a full question without requiring the reader (or the crawler) to stitch together context from elsewhere on the page. For the practical writing guide built on this research, with real before/after rewrites, see how to write content AI can quote; for the paper itself (the nine methods tested and the domain-dependent finding most summaries drop) see the GEO paper, explained.
Stage 4: do you have off-site and earned presence?
What it is: Third-party reviews, genuine press coverage, roundups and comparison pieces you didn't write, Reddit and YouTube presence you didn't pay for. Not your blog. Not your changelog. What other people, publications, and platforms say about you.
Why this is where growth actually happens, and why it's slower and more expensive than everything before it: Muck Rack's Generative Pulse team has now run three editions of its "What Is AI Reading?" study (July 2025, December 2025, and May 2026) analyzing more than 25 million cited links across ChatGPT, Claude, and Gemini. Earned media has accounted for 82–89% of all AI citations in every single edition, holding remarkably stable across nine months and three model-update cycles. Paid and advertorial content: 0.3% of citations. Journalism alone: consistently 25–27% of all links cited. This is not a blip tied to one snapshot. It's the most stable finding in the entire body of AI-search research surveyed for this piece.
Reviews specifically move fast, faster than most marketers assume. Kevin Indig's June 2026 study tracked 2,514 G2 products that had a review burst (five or more published, non-stub reviews inside a 14-day window) against a matched control group without one. Median time from the burst to the first measurable citation lift (defined as citations clearing 1.5x baseline): four days. That's a genuinely fast feedback loop for a channel most teams treat as a slow, six-month brand exercise. The same body of G2 research found the review-count effect dwarfs the paid-listing effect: going from zero to 500+ reviews lifts median citations by roughly 2,100x within the free tier alone, while a paid listing on top of an equivalent review count adds only a 1.5–2.5x multiplier once review volume and company scale are properly controlled for. Reviews are the dominant variable. Paid placement is a minor multiplier on top of it, not a replacement for it.
Why it's Stage 4 and not Stage 1: Because it's the most expensive lever to pull, and pulling it before Stage 1–3 are fixed means you're generating earned coverage that links to a site an AI crawler can't reach, describing an entity that doesn't resolve cleanly, in a format the model can't extract a clean answer from. Off-site work compounds, but only on top of a foundation that's actually load-bearing.
Stage 5: are you monitoring on an ongoing basis?
What it is: Tracking how often you're mentioned across a real prompt set, whether you're staying in the consideration set over time (not whether you appeared once), and catching factual drift before it calcifies: stale pricing a model keeps repeating, a category description that's gone out of date, a competitor's claim about you that's propagating uncorrected.
Why this is last, and why selling it first is close to malpractice: This is the stage most GEO vendors lead with, because a dashboard with a number on it is the easiest thing to sell and the easiest thing for a buyer to understand at a glance. It's also the stage that tells you the least in isolation. Track consensus on a site with a Stage 1 access problem and you're measuring absence you already knew about, using a much more expensive tool than a server-log check. Given the SparkToro consistency finding (under 1-in-100 odds of an identical brand list across runs), monitoring only becomes meaningful once you're tracking a trend across dozens of prompts and multiple models over weeks, never a single spot-check treated as a scorecard.
What "done right" looks like: A fixed prompt set (not one you improvise per session), run across multiple models on a schedule, tracking mention frequency and consideration-set persistence as a rolling trend, plus a lightweight factual-error watch, because a model repeating your 2024 pricing in 2026 is a fixable, concrete problem hiding inside an abstract-sounding "visibility" report. For the actual build (real API options, real costs, and the one platform you genuinely can't replicate this way) see how to build your own AI visibility tracking without a vendor tool.
What does everyone sell, and what does the data actually support?
| Common pitch | What the primary research actually shows |
|---|---|
| "Get your AI visibility score": a one-time audit sold as a baseline | Under 1-in-100 odds two runs of the same prompt return the same brand list (SparkToro, 2,961 runs). A single score is a snapshot of noise, not a baseline. |
| "Add schema markup and an llms.txt file for AI citations" | 97% of llms.txt files get zero requests across 137,000 domains (Ahrefs). Schema's effect on AI citations is statistically indistinguishable from zero, and negative on AI Overviews, in a matched 1,885-page study (Ahrefs). |
| "X% more AI citations in 90 days": a guaranteed lift | No independent study shows a durable citation lift on a fixed timeline, net of brand size and existing review volume. Reviews show a four-day first lift (Indig), but scaled, durable gains compound over quarters, not a single sprint. |
Why does the order matter more than any individual tactic?
Every stage above has real research behind it. None of that research helps you if you run the stages out of sequence. Off-site earned coverage pointed at a crawler-blocked site is wasted spend. A monitoring subscription on a brand with no entity resolution is measuring a hole you already knew was there. Content restructuring on pages a search engine's index doesn't contain doesn't get read by anyone, human or machine.
The five stages for winning in the AI search era aren't five independent tactics you can pick and choose from a menu. They're a dependency chain. Fix what's binary and free to verify first: access, indexation, entity hygiene. Invest in what has the best on-site evidence next: extractability. Spend real money on what compounds slowest and matters most: earned presence. Only then does measurement mean anything.
Most of the industry sells stage 5 first because it's the easiest demo. I run stage 1 first because it's the only stage where being wrong costs you everything downstream and being right costs you almost nothing to confirm.
Frequently asked questions
What are the five stages for winning in the AI search era?
Access and indexation, entity consistency, content extractability, off-site and earned presence, and ongoing monitoring, run in that fixed order, each stage gating the next.
Why does the order matter more than the individual tactics?
Because every stage depends on the one before it holding up. Off-site earned coverage pointed at a crawler-blocked site is wasted spend; a monitoring dashboard on a brand with no entity resolution is measuring a hole you already knew about. Sequencing is what turns five separately-valid tactics into a working strategy.
Do llms.txt and schema markup actually help AI citations?
Not on the evidence available as of mid-2026. Ahrefs found 97% of llms.txt files across 137,000 domains received zero AI-crawler requests in a month, and a separate matched study of 1,885 pages found schema markup's effect on AI citations statistically indistinguishable from zero, with a small negative effect on Google AI Overviews specifically.
What's the single highest-confidence lever in this framework?
Content extractability (Stage 3): it's the only stage backed by a controlled academic study (the Princeton/Georgia Tech GEO paper) rather than a vendor's self-interested dataset.
Sources referenced
- SparkToro / Gumshoe.ai AI recommendation consistency study (Jan 2026, 2,961 prompt runs)
- Ahrefs llms.txt server-log study (Jun 2026, 137,210 domains)
- SE Ranking llms.txt citation-correlation study (2026, ~300,000 domains)
- Ahrefs schema-markup difference-in-differences study (May 2026, 1,885 pages vs. 4,000 controls)
- Princeton/Georgia Tech GEO paper (arXiv 2311.09735)
- Muck Rack Generative Pulse, "What Is AI Reading?" (Jul 2025 / Dec 2025 / May 2026 editions, 25M+ links)
- Kevin Indig / G2 review-burst citation-lift study (Jun 2026, 2,514 products)
- Similarweb AI search platform-share data (Jul 2026)
- Ahrefs AI Overviews / AI Mode source-divergence study (Dec 2025)
For the shorter, standalone definition of what an AI visibility strategy actually is, see AI visibility strategy. For a concrete week-by-week timeline built on this exact sequence, see what a 90-day AI visibility roadmap actually looks like. For what "AEO/GEO/SEO" actually means term by term, see AEO vs GEO vs SEO: what's actually different and what is a generative engine, technically?
