TL;DR: There are three things you can change: whether crawlers can reach you, whether your pages can be quoted, and whether independent sources agree about what you are. Everything legitimate in this field is one of those three. Work out which one is your problem before doing any of them, because the most common failure is pulling the third lever when the first is broken.
Key Takeaways
Reachable, quotable, corroborated. Those are the three levers. Anything claiming to be a fourth is either one of these restated or something nobody can actually do.
They are not interchangeable. Each fixes a different failure, and the diagnosis takes an hour. Content work cannot fix a blocked crawler or an entity problem.
Order matters and it is fixed. Reachable first, because the other two depend on it. Corroborated last, because it is the slowest and most expensive.
Quotable is mostly structure, not quality. Good writing buried under four hundred words of preamble contributes nothing extractable.
Corroboration has to be earned. The shortcuts, arranged mentions and bought placements, breach platform rules and advertising guidance, and are detectable.
You are moving a probability, not setting a position. Doing all three well raises the chance you appear. Nothing guarantees any particular answer.
Anchor targets assume your CMS generates slug IDs from heading text. Confirm on staging and adjust the hrefs if your platform slugifies differently.
Why there are only three
Work backwards from how an answer gets made. An assistant decides whether to search, retrieves documents from an index, selects passages from them, and composes a response, usually weighing sources that agree with each other.
At each point there is exactly one thing you can influence.
Stage | What you influence | The lever |
|---|---|---|
Retrieval from an index | Whether your pages are in it and reachable | Reachable |
Passage selection | Whether a clean answer can be lifted from your page | Quotable |
Composing with weighting | Whether independent sources say the same thing about you | Corroborated |
Deciding whether to search | Nothing | Not a lever |
Which documents rank on a given run | Nothing directly | Not a lever |
That is the whole set. Every legitimate tactic in this field maps onto one of the three, and recognising which one a piece of advice belongs to is the fastest way to evaluate it.
How to use this
When you read a recommendation, ask which lever it pulls. Schema markup is reachable and quotable. Publishing a comparison table is quotable. Getting listed on a review platform is corroborated.
If a tactic does not map to any of the three, it is usually describing something nobody can do. Optimising your position in a model's vector space, correcting its training data, or timing your publishing to an indexing schedule all fail this test.
Why the framing helps
Most teams do all three at once, badly, and cannot tell what worked. Three distinct levers with three distinct diagnoses lets you fix the one that is actually broken.
It also makes the sequencing obvious, which matters because the levers depend on each other in one direction only.
What this page is not
A full explanation of each lever. Each has its own article covering the mechanism and the detail, and this page links to them rather than restating.
What it gives you is the map: what the complete set is, which one you need, and in what order.
Lever one: reachable
Whether a crawler can fetch your pages and whether the content is in what it fetches. Binary, cheap to check, and the one that invalidates everything else when it fails.
What it covers
Item | The failure |
|---|---|
robots.txt on every subdomain | A rule blocking AI crawlers that nobody decided on |
Server-side rendering | Content assembled in the browser is not in what a crawler fetches |
CDN and firewall rules | Bot management challenging crawlers regardless of robots.txt |
Login walls | Documentation and help content behind a sign-in |
Status codes served to bots | 403s and soft 404s that nobody sees in a browser |
Standard structured data | Organization, Product, FAQPage where relevant |
Why it comes first
If your pages cannot be fetched, your writing cannot be selected and your corroboration has nothing to point at. A blocked crawler makes the other two levers inoperative, not merely less effective.
It is also the cheapest to check, at about an hour, and the fix is frequently a single configuration line.
The check
Read your robots.txt on the main domain and every subdomain separately. Open a key page and view source, searching for a sentence from the body. Then filter the last thirty days of server logs by the AI crawler user agents and look at the status codes.
If robots.txt permits a crawler and the logs show nothing from it, something between the two is intervening, and that is almost always a CDN rule.
The distinction people get wrong
Google-Extended governs whether your content is used for training Gemini. It does not govern whether you appear in AI Overviews, which follows ordinary Googlebot access. GPTBot is OpenAI's training crawler, while OAI-SearchBot serves ChatGPT search.
Teams routinely block both when they intended to block only training. Read each line asking what it was meant to achieve.
What this lever cannot do
Make you appear. Being reachable is necessary and nowhere near sufficient: it gets you into the candidate pool and nothing more.
That is why teams who fix an access problem and expect a transformation are often disappointed. What they have done is stop losing automatically.
Lever two: quotable
Whether a clean, complete answer can be lifted from your page without rewriting. This is structural rather than a matter of quality, which is why well-written pages frequently fail it.
The mechanism that explains it
Retrieval selects passages, not whole pages. A page can be fetched, indexed and genuinely excellent, and still contribute nothing if no self-contained passage in it answers the question.
That single fact explains most of what follows.
What it covers
Item | The failure |
|---|---|
Answer in the first hundred words | Four hundred words of context before the page addresses its own question |
Headings that are the question | Our Approach, rather than the thing someone asked |
Self-contained paragraphs | Paragraphs that only make sense after the three preceding them |
Specifics over adjectives | Fast, powerful, intuitive, where a number or a condition belongs |
One question per page | Five questions crammed together, winning none of them |
Tables as real markup | Comparison data rendered as an image |
Dates on anything that ages | Pricing, comparisons and figures with no date |
Explicit negatives | What you do not do, which few publish and which gets quoted |
The test
Take a page and read the first hundred words. Could a stranger lift two sentences that answer the question completely, without the rest of the page?
If not, that is the fix, and it is usually a rewrite of the opening rather than the whole page.
Why marketing pages fail it
They are built to make a case before delivering a conclusion, which works for a human scrolling and badly for extraction. The instinct to build up to the answer is exactly wrong here.
Documentation, by contrast, usually passes without anyone having tried: it is plain, structured, and answers precise questions precisely.
The supporting evidence
Published research on generative engine optimisation, from a team at Princeton, Georgia Tech, the Allen Institute and IIT Delhi, found that adding citations, quotations and statistics improved source visibility in generated answers by up to around 40 percent in their tests.
That is one study on a benchmark rather than a general law, but it points the same way as everything else here: specific, attributable, checkable content outperforms assertive content.
Lever three: corroborated
Whether independent sources agree about what you are. The slowest, the most expensive, and the one with the most disreputable advice attached to it.
Two distinct parts
Consistency. Your name, category, founding details, location, leadership, pricing structure and feature set saying the same thing everywhere. Contradictions across sources make you harder to describe confidently.
Independence. Sources other than you asserting those things. A brand described only on its own website has nothing supporting it.
Consistency is mostly free and entirely within your control. Independence has to be earned, which is why it takes quarters.
What it covers
Item | Effort |
|---|---|
Identical core facts across every profile you control | An afternoon |
Organization schema with sameAs links to real profiles | An hour |
Accurate category placement on review platforms | Often a support ticket |
Complete, current review platform profiles | An afternoon |
Genuine press, analyst and podcast coverage | Quarters |
Inclusion in category roundups, pitched on merit | Quarters |
Real participation in communities, affiliation disclosed | Ongoing |
Start with consistency
The free half is also the half most often neglected. Open five places your company is described and compare the facts. Most companies find a departed founder, an old headcount, superseded pricing and a category they moved on from two years ago.
Fixing those costs an afternoon and removes contradictions that make every other source harder to weigh.
What not to do
Tactic | Problem |
|---|---|
Arranged mentions on forums presented as organic | Breaches platform rules, and FTC guidance requires disclosure of material connections |
Bought links framed as AI authority signals | Google's link spam policy names purchased links and automated link building. Renaming the tactic does not change it |
Paid placement in roundups presented as editorial | A disclosure problem, and detectable |
Review incentives | Breaches most platforms' terms |
Creating your own Wikipedia or Wikidata entry | Wikidata discourages self-promotional items and Wikimedia requires paid editors to disclose |
The risk is not only enforcement. Communities detect manufactured activity reliably, and the reputational outcome when they do is worse than the visibility was worth.
The honest version
Research worth citing, data nobody else has, tools people actually use, and genuine participation under your own name. Slow, durable, and it is the only route that does not create a liability.
Which one is your problem
An hour, and it decides where a quarter of effort goes.
The diagnostic
Take ten questions your buyers ask. Run each three times across two assistants. For each, record whether you were named and whether your site was cited. Then read the pattern.
Pattern | Diagnosis | Lever |
|---|---|---|
Never named, never cited, on anything | Not reachable, or not recognised as a distinct entity | Reachable, then corroborated |
Named sometimes, never cited | Known from third parties, nothing quotable on your site | Quotable |
Cited on some questions, absent on others | Coverage gap on the absent ones | Quotable, specific pages |
Named and cited but never recommended | Useful as a source, not seen as an option | Quotable, fit and positioning statements |
Named, but described as the wrong kind of thing | Entity problem | Corroborated, consistency half |
Competitors cited from sources you are absent from | Independence gap | Corroborated, the slow half |
The confirming check
If the pattern suggests reachable, confirm it in server logs rather than inferring. Absence from answers has several possible causes and only logs distinguish a blocked crawler from an unconvincing page.
That confirmation takes ten minutes and prevents a quarter spent writing against an access problem.
The most common real answer
Quotable, for companies with an established site and reasonable market presence. The pages exist, they are reachable, and they bury their answers.
The second most common is the consistency half of corroborated, because almost nobody maintains their external profiles after a repositioning.
When it is more than one
Usually it is. Fix in order anyway, because the order is not a preference, it is a dependency: quotable work on unreachable pages does nothing, and corroboration pointing at pages that cannot be quoted wastes the hardest-won asset you have.
Re-run it quarterly
The answer changes. A company that fixed access and rewrote its pages has a corroboration problem next, and continuing to pull the same lever past the point of return is how programmes stall.
Same ten questions, same method, once a quarter.
The order
Reachable, quotable, corroborated. Fixed, because each depends on the one before it.
The dependencies
If this is broken | Then this does nothing |
|---|---|
Reachable | Quotable. Nobody can select a passage from a page they cannot fetch |
Reachable | Corroborated. Third parties point at pages that cannot be read |
Quotable | Corroboration's return. Coverage sends systems to a page with nothing extractable |
The dependency runs one way only, which is what makes the order non-negotiable rather than a matter of taste.
What this looks like as a plan
Phase | Work | Duration |
|---|---|---|
1 | Crawler access, rendering, CDN rules, login walls | 1 to 2 weeks |
2 | Consistency: identical facts everywhere, schema, profile accuracy | 2 to 3 weeks |
3 | Rewrite the ten most commercially important pages for extraction | 4 to 6 weeks |
4 | Write new pages against identified gaps | Ongoing |
5 | Earned coverage and genuine participation | Ongoing, measured in quarters |
Consistency sits at phase 2 deliberately, ahead of content work, because it is cheap, fast and it determines whether anything you publish gets attached to the right entity.
The inversion to avoid
Starting at phase 5. It is the most visible work, it feels like marketing, and it is the slowest and most expensive. A company doing earned coverage with blocked crawlers is paying for attention it cannot capture.
This inversion is common because phase 1 is unglamorous and belongs to engineering, while phase 5 has a budget line and an agency attached.
When to run them in parallel
Once phase 1 is confirmed clean, phases 2 and 3 can run alongside each other, since different people do them. Phase 5 can start any time after phase 1, because its lead times are so long that waiting costs you a quarter.
What cannot overlap is anything before phase 1 is verified.
What no lever can do
Worth stating plainly, because a good deal of marketing in this category implies otherwise.
Claimed | Reality |
|---|---|
Edit or correct a model's training data | Training corpora are assembled by providers and are not third-party editable |
Optimise your position in a model's vector space | No external access to commercial embeddings. Nothing measures or tunes it |
Submit a correction when an assistant is wrong about you | No correction channel exists. You fix the sources and wait for retrieval |
Guarantee placement in answers | Nobody controls the output. You influence a probability |
Time publishing to an indexing schedule | Cadence is unpublished, varies by site and page, and changes without notice |
Make an assistant search when it would not | Its own decision, and your content cannot influence answers that retrieved nothing |
The pattern in the false claims
Each describes direct control over a system nobody outside the provider operates. The levers that do exist are all indirect: you change what is available to be retrieved, and the system does what it does with it.
That is an uncomfortable thing to sell, which is why the false versions circulate.
What the honest version sounds like
You are raising the probability that you appear, not setting a position. Doing all three levers well makes you a much more likely candidate and guarantees no particular answer on any particular day.
That framing also sets correct expectations about measurement: you judge this on rates across many questions over quarters, never on whether you appeared in one answer this morning.
The practical consequence
When something you cannot explain happens, the honest answer is sometimes that the system changed and nobody said so. Recording competitors alongside yourself is what lets you tell that apart from something you did.
If everyone moved, it was the system. If only you moved, look at your own site first.
How long each takes
The three levers operate on completely different timescales, which is the main reason expectations go wrong.
Lever | Effort | Time to effect |
|---|---|---|
Reachable | Hours to days | Days to weeks once re-crawled |
Quotable | Weeks | Weeks to a quarter per page |
Corroborated, consistency | An afternoon | Weeks to months, as sources are re-read |
Corroborated, independence | Ongoing | Quarters, sometimes a year |
Why the spread matters
A team that fixes all three in month one and expects results in month two is disappointed by the third, which has barely started. A team that reports honestly says access is fixed and already showing, pages are published and we are watching, coverage is in progress and will not show for two quarters.
Setting that out at the start is what keeps a programme funded through the quiet middle.
What moves fastest
Access fixes, by a wide margin. If crawlers were blocked and now are not, answers can change within weeks, and it is the only lever that produces a fast visible result.
That makes it useful politically as well as technically: it is the early evidence that the work does something.
What moves slowest
Category position in a competitive market. Being recommended where three established names currently dominate is a quarters-to-years objective, and no amount of spending compresses it much.
If that is your goal, say so with the timeline attached, because the gap between expectation and reality is what kills these programmes in month three.
Measuring per lever
Lever | The number that moves |
|---|---|
Reachable | Crawler requests in server logs, then citation rate |
Quotable | Citation rate, and specific prompts flipping to cited |
Corroborated | Mention rate, and whether you are described correctly |
Tracking them separately tells you which lever is paying, and that is what you need before deciding where the next quarter goes.
Six mistakes
1. Pulling lever three first. Earned coverage is the most visible work and the slowest. Doing it while crawlers are blocked pays for attention you cannot capture.
2. Treating quotable as a writing-quality problem. It is structural. An excellent page that buries its answer in the sixth paragraph contributes nothing extractable, and the fix is the opening rather than the prose.
3. Skipping the consistency half of corroboration. It is free, it takes an afternoon, and almost nobody does it after a repositioning. Contradictory facts across sources undermine everything else.
4. Buying corroboration. Arranged mentions and purchased links breach platform rules and run into advertising guidance and Google's link spam policy. They are detectable, and the reputational cost when detected exceeds the visibility.
5. Believing in a fourth lever. Vector positioning, training data correction, indexing-schedule timing. Each describes direct control over a system you do not operate, and no tactic built on them can work.
6. Not diagnosing before acting. An hour of running ten questions tells you which lever is broken. Teams that skip it usually pull the lever they are most comfortable with, which is rarely the broken one.
The pattern
Four of six come from doing the work that feels like marketing rather than the work that is blocking you. Access and consistency are unglamorous, belong to other teams, and are where most of the available gain sits.
The one that wastes most
Number six, because it causes the other five. An hour of diagnosis is the cheapest thing in this article and it decides where a quarter goes.
Ten questions, three runs each, two assistants. Then read the pattern and pull the lever the pattern names rather than the one you were already planning to pull.
FAQs
How do you optimize for AI search engines?
Three things, in order. Make your pages reachable by crawlers, make them quotable by answering one question directly in the opening, and make independent sources agree about what you are. Everything legitimate in this field is one of those three.
Which should I do first?
Reachable, always. Check robots.txt on every subdomain, confirm pages are server-rendered, and look for AI crawler hits in server logs. If crawlers cannot fetch your pages, content work and earned coverage both do nothing.
How do I know which lever is my problem?
Run ten buyer questions three times across two assistants and read the pattern. Named but never cited points at quotable. Never named at all points at reachable or entity recognition. Described as the wrong kind of thing points at consistency.
What makes a page quotable?
A complete answer in the first hundred words, a heading that is the question, self-contained paragraphs, specifics instead of adjectives, one question per page, and real tables rather than images. Retrieval selects passages, so a buried answer contributes nothing.
Does schema markup help?
Yes, modestly. Organization, Product and FAQPage markup where genuinely relevant helps systems parse what your pages describe. It is a contributor rather than a lever on its own, and FAQPage markup must match visible page text.
How important are third-party sources?
Important, and slow. A brand described only on its own site has nothing corroborating it. Start with the free half, which is making your facts identical everywhere you control, then earn independent coverage over quarters.
Can I pay for mentions to speed this up?
No. Arranged mentions presented as organic breach platform rules and run into FTC guidance on disclosing material connections, and bought links are named in Google's link spam policy. Both are detectable and the reputational cost exceeds the gain.
How long before any of this works?
Access fixes can change answers within weeks. Page rewrites take weeks to a quarter each. Earned corroboration takes quarters, sometimes a year. Set those three expectations separately or the programme gets judged on the slowest one.
Where to start
Spend the hour on diagnosis. Ten questions your buyers ask, three runs each, two assistants, recording whether you were named and whether you were cited. The pattern tells you which lever is broken.
Most teams skip this and pull the lever they are most comfortable with, which is usually the third and usually not the problem.
The three levers
Reachable. robots.txt on every subdomain, server-side rendering, CDN rules, login walls, crawler hits in logs.
Quotable. The answer in the first hundred words, the question as the heading, one question per page, specifics over adjectives.
Corroborated. Identical facts everywhere you control first, then earned independent coverage over quarters.
In that order, because each depends on the one before it.
If the bottleneck is writing
Lever two is a volume problem once you know which pages to fix: a list of specific questions, each needing a page that answers it directly. Producing those consistently is where most teams stall, and Blazly SEO covers research, planning, drafting and direct publishing from about $39 a month.
It does not fix lever one, which is an engineering job, and it does not earn you lever three, which has to be genuine. It helps with the middle one, which is where most of the ongoing work sits.
Summary
There are exactly three things you can change about how AI search treats you. Whether crawlers can reach your pages, which is binary, cheap to check and invalidates everything else when broken. Whether a clean answer can be lifted from your pages, which is structural rather than a matter of writing quality, and which most marketing pages fail because they build to a conclusion. And whether independent sources agree about what you are, which splits into a free consistency half you should do immediately and an earned independence half measured in quarters. Pull them in that order, because the dependencies run one way. Diagnose first, since an hour of running questions tells you which one is broken and prevents a quarter spent on the wrong one. And remember that all three move a probability rather than setting a position, so judge the work on rates across many questions over quarters, never on a single answer.