The Complete Guide to AI SEO

Chapter 1

The new retrieval landscape

Search did not disappear. It gained a layer. Understanding that layer is the difference between adapting and panicking.

The surfaces that now matter

AI Overviews sit above organic results on a growing share of queries. When present, the traditional first result drops below the fold on most phones.

Assistants are used for vendor research, shortlisting and comparison. These are commercial queries that previously produced the most valuable clicks.

In-product search inside the tools your buyers already use, increasingly answering rather than listing.

How each surface gets its information

Not all of them work the same way, and the differences change what you do.

The practical takeaway: most of these read the live web through a crawler. Access, speed and structure decide whether you are in the running.

What actually changed for practitioners

Three shifts:

1. The unit of competition moved from page to passage. Engines retrieve and evaluate chunks, not documents.
2. Citation replaced position. Being named in the answer is the visibility.
3. Corroboration became a ranking input. A claim repeated across independent sources becomes a fact worth quoting.

What did not change

Almost everything else. Crawlability, relevance, clarity, authority and originality all still matter, and matter more.

The teams who do well here are usually the ones who were already doing search properly. The tactics that stop working are the ones that were always gaming: keyword density, thin location pages, content padded to a word count.

Chapter 2

Being chosen: the source model

Three filters decide whether you appear. A page must pass all three, and most effort goes to the last while failing the first.

Filter one: retrieval

Mechanical and binary.

• Crawler access. Check robots.txt for GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot, Google-Extended. Also check your CDN or WAF, which can block independently.
• Rendering. If content requires JavaScript, expect partial or empty extraction.
• Response speed. Live retrieval systems time out.
• Index presence. Not indexed means not a candidate.

Test this before anything else. It takes an hour and it is where most of the loss occurs.

Filter two: passage relevance

Engines chunk documents and score chunks.

The consequence is counterintuitive: a 700 word page answering directly can beat a 3,000 word guide that buries the answer in paragraph fourteen. The long guide may be the better document. The short one is the better chunk.

This is why answer-first structure is mechanical rather than stylistic.

Filter three: corroboration and trust

Among relevant passages, engines prefer verifiable claims from publishers with reason to be trusted.

Signals that demonstrably matter:

• Consistent business identity across the web
• Named authors with real credentials
• Specific, checkable claims
• Agreement with independent sources
• Original data that others reference

The diagnostic

Run this before changing anything.

The strategic implication

Once filter three is understood, a durable approach emerges: publish something specific and true that nobody else has, make it easy to quote, and get it referenced.

At that point you stop competing for a citation slot and become the source of the fact.

Chapter 3

Content for answers

The answer-first structure

Every page intended to earn citations follows the same skeleton.

1. Direct answer, first 60 words. Answer the title question plainly. Test: could this paragraph be quoted with no other context?

2. Supporting context, 150 to 300 words. Conditions, exceptions, reasoning. This is what stops the direct answer from being glib.

3. Structured detail. Tables, numbered processes, comparisons. Extracted with high fidelity and attributed accurately.

4. Related questions, each with its own subheading and immediate answer. The highest-leverage section, because it multiplies the queries one page can serve.

5. Provenance. Author, credentials, last verified date.

Rewriting an opening

Before: "In today's rapidly evolving digital landscape, businesses face unprecedented challenges when selecting a marketing partner."

After: "A competent SEO retainer for a mid-market Indian business runs fifty thousand to two lakh rupees monthly, scaling with market competitiveness rather than site size. Below thirty thousand you are buying part-time junior time, not strategy."

The second can be quoted. The first cannot be quoted by anyone, for any purpose.

Question architecture

Build around question clusters rather than keywords.

Comparative content is where most brands lose nerve. A comparison concluding you win everything is discounted immediately. One that names where a competitor is the better choice gets cited, because it reads as evaluation rather than marketing.

Formatting that measurably helps

• One idea per paragraph
• Subheadings phrased as questions where natural
• Numbers over adjectives
• Define terms on first use
• Conclusion before reasoning
• One clear topic per URL

What to stop

• Keyword density work, which is noise to semantic retrieval
• Padding to word counts, which dilutes chunk relevance
• Thin location pages that corroborate nothing
• Anonymous publishing
• Hedged answers containing no specifics

Chapter 4

Entity foundations and corroboration

Content decides whether you are quotable. Entity work decides whether the engine knows who is being quoted.

Consistent identity

The same name, contact details and description everywhere you appear. Inconsistency prevents an entity from consolidating, because the system cannot tell whether it is looking at one company or three.

Audit by searching your brand name and listing every property describing you. Fix the disagreements.

Organization schema, properly complete

Most sites ship a stub. A useful block declares:

• Legal name and trading names
• Logo as a proper ImageObject with dimensions
• Contact points with type and language
• Areas served
• Fields of expertise
• Links to every profile you control

That last field is a direct statement that these scattered profiles are one entity.

Page-level structured data

Fabricated markup is worth calling out. Review schema without reviews is a manual action risk and, more immediately, it teaches the engine your structured data is unreliable.

Author identity

Named authors with a real page: credentials, history, external work. The weighting gap between authored and anonymous content is substantial, and largest for anything touching money, health or law.

Building corroboration

Schema is your claim about yourself. Corroboration is others making the same claim.

In order of effort:

1. Complete, consistent directory profiles
2. Association and partner listings
3. Review platforms relevant to your category
4. Guest bylines under a named person
5. Original research others cite by name

The fifth compounds and the others do not. A published benchmark that becomes the number people quote makes you the source rather than a candidate.

llms.txt

A plain-text summary at your root, written for machine consumption. Support is not universal and it is not a ranking factor.

It is also nearly free, harmless if ignored, and a clean statement of structure for systems that read it. Publish one, generate it from live content so it cannot drift, and do not expect it to substitute for content and entity work.

Chapter 5

Measurement and iteration

What you cannot measure

State the limits before promising anything internally.

• No impression counts for answer engines
• Answers vary by user, session and phrasing
• No citation API for arbitrary brands
• Attribution from answer to conversion is indirect

Anyone promising exact AI impression share is selling a model, not a measurement.

What you can measure

Citation share by prompt set

1. Build 40 to 60 prompts a real buyer would ask, across definitional, comparative and evaluative intents
2. Run monthly across relevant engines, clean session, no personalisation
3. Record: cited or not, which URL, which competitors appeared
4. Track the percentage where you appear

Freeze the prompt set. Changing it monthly destroys comparability, which is the whole point.

Assistant referral traffic

Identifiable referrers in analytics. Segment them. Volume looks small, quality is usually high, because the visitor arrived after reading an evaluation.

Judge on conversion rate, not volume. Comparing to organic session counts is the wrong comparison and will always look bad.

Crawler access logs

The most underused data in the discipline. Server logs show which agents fetch which pages and what they receive.

• Are crawlers reaching priority pages
• Are they getting 200s or errors
• Which content attracts repeat retrieval
• Did access change after a config edit

Branded demand

The lagging commercial indicator. Working AI visibility raises branded search and direct traffic.

Monthly report

Then one sentence: what changes next month, and why.

Realistic timelines

• Weeks 1 to 4: access and technical fixes. Citation share barely moves.
• Weeks 4 to 12: restructured content indexed. Definitional prompts move first.
• Months 3 to 6: comparative and evaluative prompts shift. This is the commercially valuable set and it moves last.
• Months 6 plus: entity and corroboration work compounds.

Do not read early definitional wins as proof that commercial prompts will follow automatically. They require the corroboration work.