AI Search Ranking Factors: What LLMs Actually Reward

2. Why “Ranking Factors” Is a Borrowed Word Here

Quick Answer: AI answers are not ranked lists. Google describes its generative features as retrieval-augmented generation sitting on top of ordinary Search ranking, plus query fan-out into related sub-questions. So AI search ranking factors are mostly search ranking factors, applied to a passage rather than a page.

Google’s guide to generative AI features spells out two mechanisms. The first is retrieval-augmented generation, which relies on core Search ranking systems to fetch pages from the Search index. The second is query fan-out: the model quietly generates related queries around yours and pulls results for those too.

Both matter for planning. Retrieval means an unindexed page cannot be cited at all, which makes deciding which AI crawlers to allow through a prerequisite, not an optimisation. Fan-out means the winning page is often the one answering a sub-question nobody wrote a page for.

So there is no separate algorithm to reverse-engineer. There is retrieval you must qualify for, and selection you must earn at passage level. That distinction underpins optimising a site for large language models more usefully than any numbered list.

3. The Three Gates: Retrieved, Quoted, Recommended

Quick Answer: Three separate things have to happen before an engine recommends you: your page must be retrieved, a passage from it must be worth quoting, and the recommendation itself must be supported by evidence off your own site. Different fixes clear different gates.

Most lists blur these together, which is why the advice contradicts itself. Schema cannot help a page nobody retrieves. Backlinks cannot help a page with no quotable sentence in it. Sort the work by gate:

• Gate one, retrieval. Indexed, crawlable, not blocked. Binary, cheap to fix, and the only gate where technical work alone is decisive.
• Gate two, quotation. A self-contained passage somewhere on the page answers the question without the paragraphs around it. Where most Malaysian SME sites fail.
• Gate three, recommendation. The engine puts you forward as an option, not just a source. That almost always needs corroboration from somebody else.

Knowing your gate changes the shopping list. Establishing it is the purpose of an AI visibility audit, and it explains why referral counts mislead: the quality of AI-referred visitors depends on which gate produced the mention.

4. What Separates a Quoted Page From an Ignored One

Quick Answer: Comparing pages that earned a citation against pages on the same sites that never did, the widest gaps are a direct answer near the top, a specific figure or date, and a named author with verifiable expertise. Schema markup shows a much narrower gap.

The comparison holds the site constant, so domain strength is not doing the work. Every trait is one an editor can add in an afternoon.

The last row is the interesting one. Schema is nearly as common on ignored pages as on cited ones, so it explains little about which page got quoted — a hygiene item rather than an AI search ranking factor worth budgeting against.

5. Five AI Search Ranking Factors Google Says Do Nothing

Quick Answer: Google’s published guidance explicitly rules out llms.txt files, content chunking, rewriting pages for machines, chasing inauthentic mentions, and treating structured data as a requirement. All five appear on widely-shared AI search ranking factor lists.

This is the rare case of a platform saying the quiet part out loud. Google’s mythbusting section on generative AI search lists what you can ignore:

• llms.txt and other special files. Google says plainly that Search does not use them, and keeping one neither helps nor harms — our conclusion too on whether your site needs an llms.txt file.
• Chunking your content into fragments. No requirement to break pages into tiny pieces, and no ideal page length.
• Rewriting content for machines. The systems understand synonyms and intent, so covering every phrasing variant is wasted effort.
• Buying or manufacturing mentions. Inauthentic mentions are described as less helpful than they look, and spam systems act on them.
• Structured data as a requirement. No special schema is needed, though it stays worth keeping for rich results — the view we take in what schema to add first.

Read that list against a typical agency proposal and a lot of line items thin out. Ask any provider selling generative engine optimisation services which of the five they charge for. The better answers come from answer engine optimisation specialists who can explain what they deliberately do not sell.

6. Where the Engines Actually Pull Their Sources From

Quick Answer: Across tracked Malaysian buying prompts, company websites supply well under half the cited sources. Review roundups, forum threads and directories together supply more, which is why on-page work alone rarely lifts a business into a recommendation.

This distribution explains more disappointing campaigns than any other number we track.

So getting listed accurately in the roundups and directories your buyers read is a citation strategy, not vanity. The forum share explains why Reddit and Quora threads keep getting quoted ahead of polished company pages. The mix shifts by engine, so check how Gemini assembles results and how Copilot builds answers.

7. Passage Extractability: The On-Page Factor That Still Moves

Quick Answer: Engines quote passages, not pages. A passage is quotable when it answers one question completely in about forty to sixty words, names its subject rather than saying “we”, and survives being read alone with no surrounding context.

This is not chunking, which Google says you do not need. Chunking fragments a page for machines. Extractability means one clean answer per question inside a page that still reads normally. Five steps get most pages there:

1. Put the answer above the argument. Lead with the conclusion, then justify it. Engines rarely read far enough to find a buried verdict.
2. Name the subject in the sentence. “ZenWeb builds Google Ads campaigns for Malaysian SMEs” is quotable; “we do this for our clients” is not — it loses meaning once lifted.
3. Keep one question per heading. Sections answering two get skipped for sections answering one cleanly.
4. Attach one concrete detail. A price band, a timeframe, a city, a figure. Detail is what wins the quote over a competitor’s identical claim.
5. Read the passage alone. If it stops making sense without the paragraph above, rewrite it until it stands on its own.

On a small site this is an afternoon’s editing. On a large catalogue it becomes a templating problem, where enterprise SEO across thousands of pages earns its keep: fix the template once, not the pages one by one.

8. Off-Site Corroboration: The Gate Most Lists Underweight

Quick Answer: Being recommended rather than merely cited depends on your business being described consistently on sources you do not control. Consistent naming, a resolvable business identity, and genuine third-party coverage do more here than anything on your own pages.

Google’s warning against manufactured mentions is worth taking literally. The distinction is between corroboration you earn and volume you buy. Four things carry weight:

• One consistent business identity. Same legal name, address format and service wording everywhere. Contradictions make an engine hedge rather than recommend, the problem entity SEO solves.
• A resolvable public record. Machine-readable facts about the company — the practical argument for earning a Google Knowledge Panel and for tidy Wikidata entries where they apply.
• Real coverage by other people. Trade press, customer write-ups, genuine forum answers. Slow, and the only version that survives spam filtering.
• Being named without being linked. Engines pick up plain-text company names, so unlinked coverage counts — the mechanism behind getting named rather than merely linked.

None of this is fast. It is also the only layer that reliably turns a citation into a recommendation, and usually the one neglected while the homepage got its third rewrite.

9. AI Search Ranking Factors Ranked by Effort Against Movement

Quick Answer: Sorting the commonly-cited AI search ranking factors by implementation effort against observed movement in cited-prompt share puts passage rewriting and listing accuracy at the top. Schema clean-up and file-based tricks sit at the bottom.

Effort is in working days for a typical Malaysian SME site. Movement is the median change in tracked prompts citing the business, six months on.

The footnote matters. Businesses rarely change one thing at a time, so read this as a priority order, not a clean attribution.

10. How Long Each Fix Takes to Show Up

Quick Answer: Technical unblocking shows up within a fortnight because it only needs a recrawl. Passage rewrites take roughly a month, new pages closer to two, and off-site corroboration a quarter or more. Expecting them on the same schedule is what makes programmes look stalled.

Sequencing matters as much as choice. Start the slow work early and let the fast work buy evidence.

Read the bottom row before signing a three-month contract. Corroboration rarely surfaces by the quarterly review, so start it first and judge it last — the ordering behind a sensible 90-day AI search roadmap.

11. How to Test Any AI Search Ranking Factor Yourself

Quick Answer: Freeze a set of buying prompts, record who gets named today, change one thing, wait a full crawl cycle, then re-run the identical prompts. Any AI search ranking factor that survives that test on your own site is worth keeping.

You do not need a tool subscription, and doing it once teaches you more than any list:

• Write twenty prompts in customer wording. Real sentences with constraints: a budget, a city, an industry. Not keyword strings.
• Record the baseline first. Note whether you are absent, named, cited or recommended in each answer, and date it.
• Change one variable. Rewrite five pages, or correct your listings, but not both in the same fortnight.
• Re-run the identical prompts. Same wording, same day of the month. Editing prompts to chase a better number ruins the comparison.
• Check the traffic side too. Whether AI-referred visitors convert is a different question from whether you were named.

Google’s own reporting has improved. AI feature appearances were once folded into overall Search Console web data, and Google has since added a dedicated generative AI performance report. That is a first-party check on one engine, pairing well with ranking in Google AI Overviews from Malaysia.

12. Conclusion

Quick Answer: The useful AI search ranking factors are unglamorous — be retrievable, write passages worth lifting, and be described consistently by other people. The rest of the published lists are either hygiene items or things Google has said it ignores.

The vocabulary will keep churning. The job will not. Engines fetch pages ordinary search already ranks, pick the passage answering the question cleanest, then lean on outside corroboration before putting a name forward.

ZenWeb has run search work for Malaysian businesses since 2000. We fold AI visibility into the same SEO service rather than pricing it separately, because the page that earns a citation is the page that earns a ranking.

13. Frequently Asked Questions

Are AI search ranking factors published anywhere official?

No. No engine publishes a weighted list. Google’s guidance on how its generative features work is the closest thing available, and it stops well short of a ranking-factor list. Treat any numbered list as inference, not documentation.

Does schema markup help me get cited by AI?

Only indirectly. Google states that structured data is not required for its generative features, and in our comparison cited pages carried schema only slightly more often than ignored ones. Keep it for rich results and SEO hygiene, but do not expect it to win a citation.

Why does my competitor get recommended when my pages are better?

Usually because the recommendation is not coming from anybody’s website. Roundups, forums and directories supply a large share of cited sources, so a competitor described consistently across them gets put forward ahead of a business with better pages and thinner outside coverage.

How long before changes show up in AI answers?

Technical unblocking typically shows within a fortnight, passage rewrites within a month, new pages closer to two, and off-site corroboration a quarter or more. Answers also shift on their own week to week, so judge movement over months.