Being visible in AI search is not one lever, and the list of things that work is shorter and less exciting than the list of things being sold. This is what actually moves it, in the order I work through it, which is roughly cheapest and most certain first.
The measurement that tells you whether any of it worked is a separate job, covered in how to track brand mentions in AI search. Do not start here. Baseline first, or you will have no way to tell an improvement from a good week.
First, be reachable
Everything below assumes the engine can fetch your pages. A meaningful share of sites with zero AI visibility have an access problem rather than a content problem, and it is usually accidental: a plugin default, a firewall rule, a robots.txt copied from a template that blocks bots somebody once read about.
Check per vendor, because they run more than one agent and the agents do different jobs. Blocking the one that trains a model while leaving the one that fetches a page for a live answer is a reasonable position; blocking both by accident is not. The specifics are in the posts on ChatGPT, Claude, Perplexity and Google-Extended.
This is a half-hour job with a yes or no answer, and no amount of writing fixes it if the answer is no.
Then, be liftable
An answer engine does not recommend a page. It takes a passage and rewrites it. So the unit of optimization is the passage, and the question for every section you write is whether a machine could lift eighty words from it and have a complete answer.
In practice that means four habits.
Ask the question as the heading. Not “Indexing considerations” but “Why is my page crawled and not indexed”. Headings that match how the question is asked are the cheapest retrieval signal available.
Answer in the first two sentences under it. Before the context, before the caveats, before the paragraph about why this matters. Put the answer first and expand afterwards. This is the single largest change most sites can make and it costs nothing but the reflex to warm up.
One idea per paragraph. A passage carrying three ideas cannot be lifted cleanly, so it will not be.
Keep the sentences self-contained. A sentence that begins “As mentioned above” is unusable out of context, and out of context is where it will be used.
Applied across an existing site this is an editing pass, not a rewrite, and it is the fastest thing on this page to do.
Then, be worth quoting
Machines do not repeat text they could have generated themselves. This is the part that separates sites that get cited from sites that merely get crawled, and it is unforgiving.
Generic advice is unquotable because a thousand pages say it equally well and the model already holds it. What gets used is the sentence carrying something specific: a number, a date, a version, a threshold, a named limit, a result you observed. “The export caps at 1,000 rows” is quotable. “Make sure to check your data” is not.
If a page contains nothing a model could not have written without you, it will not be cited, and it should not be. The practical instruction is to put at least one specific, checkable, ownable fact into every page you publish, and to prefer the things only you know: what happened on the jobs you ran, what the tool actually returned, how long it actually took.
This is also why publishing more is the wrong strategy here. Twenty thin pages give an engine twenty things not worth repeating. One page with a real finding beats the lot, and a site that pivots to volume loses the property that made it citable in the first place.
Then, be safe to attribute
A model choosing whether to name a source is, in effect, deciding whether pointing at you is defensible. Things that help: a named author who exists elsewhere, claims tied to evidence rather than assertion, dates on the material, consistent identity across your site and off it, and a clear account of who you are and what you do.
This is slower than everything above and it compounds. It is also the difference between the quoted-but-not-credited case and the cited case, which is the most common upgrade available to a site that has already done the writing work.
Being talked about elsewhere by name belongs in this bucket too. Not link building in the old mechanical sense: being referred to in the places a model has read, which includes discussions, comparisons and roundups where a link may never appear.
Then, the classic SEO you were already doing
For the Google surfaces, ranking is close to a prerequisite, because AI Overviews and AI Mode retrieve from the same index. A page that is not indexed cannot be summarised from it, and a page on the fourth page of results is rarely in the retrieval set.
So the ordinary work still applies and now has a second payoff. If pages are missing from the index, that is your bottleneck and nothing on this page will move until it is fixed: start at why pages do not show on Google, and the paid version of that work is indexing.
What does not work
Worth stating plainly, because each of these is being sold.
Publishing more. Covered above. Volume is the reflex answer and it actively hurts here.
llms.txt as a lever. It costs an hour, it is cheap insurance, and no engine is obliged to read it. Write it, keep it in sync, expect nothing. What changed in v2 has the detail.
Keyword density, in its new clothes. Repeating a phrase to make a model say it back is the 2010 mistake with a new name.
AI-written content at scale to feed AI search. The output is by construction the average of what already exists, which is precisely the thing an engine has no reason to quote.
Anything sold with a guarantee. Nobody controls what an answer engine says. A guaranteed placement is either a misunderstanding or a trick.
The order matters more than the list
Most of the value sits in the first two sections. Access is binary and cheap. Liftability is an editing pass over pages you already have. Both can be done this month, and both are measurable against the baseline you took first.
Specificity and attribution are the long game, and they are what stops the gains from being copied by the next site to read the same advice.
Then re-run the sheet, monthly, for six months, and let the numbers rather than the vendors tell you what worked. If you would rather have that run for you, it is AI search optimization, and until this site’s own number moves that page describes the method and promises no outcome.
Quick answers
How long does it take to see a change in AI visibility?
Longer than a ranking change and with a less clean signal, because you are waiting on two separate clocks. Retrieval-based answers can reflect a rewritten page within days of it being recrawled, so a fixed access problem or a restructured answer can show up in one monthly run. Anything that depends on the model itself, or on other sites referring to you, moves on a scale of months and you should not expect to attribute it to a single change. Measure monthly, judge on three runs, and be suspicious of any account of a two-week transformation.
Does schema markup improve AI visibility?
It helps at the margin and it is not the lever people hope for. Structured data makes facts unambiguous, which is useful when a machine is deciding what your page asserts, and it is cheap to add correctly. But an answer engine lifting a passage is reading the prose, not the JSON-LD, so schema on a page that buries its answer changes nothing. Do it because it is good hygiene for the classic surfaces, not as an AI strategy.
Do backlinks still matter for being cited by AI?
Indirectly, and less mechanically than in classic SEO. For the Google surfaces they feed the ranking that gets you retrieved in the first place, so they matter by proxy. Beyond that, what appears to count is being referred to by name in places a model has read, which is a broader thing than a link: a mention in a discussion, a roundup, a comparison, a forum answer. A link is one way to earn that and not the only one.
Is there any point optimizing if my competitors are much bigger?
Yes, and this is one of the few places where size helps less than usual. Answer engines are choosing a sentence to reuse, not a brand to endorse, so a small site with the exact specific answer routinely beats a large one with a general page. Where scale wins is category questions, where the model repeats the names it has read most often. Aim at the problem questions first: they are winnable, and they are what buyers actually type when they have a problem.
Should I write pages aimed specifically at AI engines?
Write the page for the person and structure it so a machine can lift from it. Those two things are almost entirely compatible, because the structure that helps a model, which is a direct answer high up under a heading that matches the question, is also what a reader in a hurry wants. What does not work is content written for machines only. It reads as filler to people, and models are increasingly trained on signals that punish exactly that.