Ask ChatGPT, Gemini or Perplexity to recommend a Razorpay integration partner in Bengaluru and you get three or four names with a couple of sentences each. No ten blue links, no page two. For the companies named it is the best lead source they have. For everyone else it is invisible — you cannot even tell you lost.
That is the shift worth planning for. A growing share of commercial research now ends inside an assistant's answer rather than on a results page, and the assistant quotes a handful of sources instead of listing everything it found. Generative engine optimization — GEO — is the work of being one of those sources. It is less a new discipline than an old one with the scoring function changed.
What is generative engine optimization?
GEO is the practice of making your content easy for a large language model to find, trust and quote when it assembles an answer. The page still has to be crawlable and useful — none of the SEO fundamentals go away. What changes is the target: instead of competing for a rank position, you are competing to be the sentence the model lifts.
You will also see this called AEO (answer engine optimization) or "AI SEO". The labels are marketing; the work is the same. What matters is understanding how the two systems differ, because the tactics that win a rank position are not all the tactics that win a citation.
| Classic SEO | GEO | |
|---|---|---|
| Unit of success | A ranked URL | A cited sentence |
| First reader | A person scanning titles | A model assembling an answer |
| What wins | Authority and keyword coverage | Extractable, verifiable claims |
| Where you see it | Search Console impressions | Referral traffic and manual prompting |
| Feedback loop | Days to weeks | Weeks, and noisy |
Why an assistant cites one page and ignores another
Retrieval systems chunk a page, embed the chunks, and pull back the few that best match the question. Then the model decides which of those chunks it can safely repeat. Both stages punish the same thing: prose that only makes sense if you have read the paragraph above it.
- Self-contained answers. A chunk that begins "This is why it matters" is useless out of context. One that begins "A working web MVP in India costs ₹2,00,000 to ₹8,00,000" can be quoted as-is.
- A question-shaped heading. "What does it cost?" retrieves against a user's phrasing. "Investment" does not.
- Specific, checkable facts. Numbers, versions, dates and named tools give a model something concrete to attribute. Adjectives give it nothing.
- A visible date. Assistants weight freshness heavily on anything with a year in the query. An undated page loses to a dated one that says less.
- Server-rendered HTML. Most crawlers used for retrieval do not run your JavaScript. If the text only exists after hydration, it does not exist.
- Corroboration elsewhere. A claim that appears only on your own site is one a model is reluctant to repeat. Being described the same way on LinkedIn, a directory and a customer's case study makes the claim safer to use.
Seven things to fix on your own pages
In rough order of effort-to-payoff. None of this needs a new platform; all of it is work on pages you already have.
- 1
Answer the title in the first 40 words
Put one self-contained sentence directly under the H1 that answers the question the title asks. This is the sentence that gets quoted. Write it so it still makes sense pasted into a chat window with no surrounding page.
- 2
Rewrite headings as questions
Turn "Pricing" into "What does it cost to build an MVP in India?". Turn "Process" into "How long does a build take?". Your H2s become the retrieval handles, and as a side effect the page gets a usable on-page index.
- 3
Add an FAQ block with the questions people actually type
Pull them from your own inbox, your sales calls, and the People Also Ask box. Three to five per page, each answered in two or three sentences. Mark it up as FAQPage so the structure is machine-readable as well as visible.
- 4
Put the numbers in
Ranges, timelines, versions, stack names. "We are experienced" is unquotable. "We have shipped 40+ Razorpay integrations since 2022, typically in 3–5 days" is a citation waiting to happen — provided it is true.
- 5
Ship Organization, Article and Breadcrumb JSON-LD
This is the cheapest structured signal available and most sites still get it wrong or skip it. It tells a retrieval system who published a page, when, and where it sits in the site.
- 6
Make every claim checkable from one page
Link the case study, name the client if you are allowed to, date the engagement. A model weighs a claim it can trace far more heavily than one it cannot.
- 7
Keep the date honest
Show publishedAt and updatedAt, and only bump the update date when you genuinely changed something. Rolling the date forward on unchanged content is the oldest trick in SEO and it is a poor bet against systems that can diff your text.
What llms.txt is, and what it is not
llms.txt is a proposal from 2024: a markdown file at the root of your domain that gives a model a curated map of your site, so it does not have to infer your structure from navigation and sitemaps. It looks like this:
# CruxBit
> A software consultancy in Bengaluru building web platforms, mobile apps and
> AI products, and running a project-based internship programme.
## Services
- [Services overview](https://cruxbit.tech/services): what we build and how we scope it
- [Pricing](https://cruxbit.tech/pricing): indicative ranges by project type
## Guides
- [MVP cost in India](https://cruxbit.tech/blog/mvp-cost-india-2026): rupee ranges by build type
- [What AI really costs](https://cruxbit.tech/blog/what-ai-really-costs-in-2026): token and infra math
## Contact
- [Start a conversation](https://cruxbit.tech/#contact)Structured data is still the cheapest win available
Schema.org JSON-LD is unglamorous and it is the single highest-leverage thing most Indian SMB sites are missing. It states, unambiguously, facts that a model would otherwise have to guess at: who you are, what you sell, where you operate, when a page was written. Here is a minimal Organization block worth having on every page:
{
"@context": "https://schema.org",
"@type": "Organization",
"name": "CruxBit",
"url": "https://cruxbit.tech",
"logo": "https://cruxbit.tech/logo.png",
"areaServed": "IN",
"knowsAbout": ["AI product development", "Next.js", "Razorpay integration"],
"sameAs": [
"https://www.linkedin.com/company/cruxbit",
"https://github.com/cruxbit"
]
}How do you tell whether any of this is working?
This is the genuinely hard part, and anyone promising you a clean GEO dashboard is overselling. There is no Search Console for assistant citations. What you can do is triangulate.
- 1
Segment assistant referrals in analytics
Traffic from chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com arrives with its own referrer. Make it a saved segment. It is small, it converts unusually well, and its trend line is the closest thing to a real metric you have.
- 2
Run a fixed prompt panel every month
Write down 15–20 prompts a genuine buyer would type — "best Next.js development agency in India", "how much does an AI chatbot cost in India" — and run them across two or three assistants on the same day each month. Record who gets named. It is manual, it is noisy, and it is still the only direct read you get.
- 3
Watch your server logs for AI crawlers
GPTBot, ClaudeBot, PerplexityBot and Google-Extended identify themselves in the user agent. If they are not fetching a page, that page cannot be cited — and that is a robots.txt or rendering problem you can fix today.
Frequently asked questions about GEO
Is GEO replacing SEO?
No. Classic search is still where most commercial traffic comes from, and the crawlability, speed and content-quality work is shared between the two. Treat GEO as an additional scoring function on the same site, not a separate project with its own budget.
How long does GEO take to show results?
Slower and noisier than SEO. Assistants refresh their retrieval indexes on their own schedule, and answers vary between runs of the same prompt. Expect to need two to three months of monthly prompt tests before a trend is distinguishable from noise.
Do I need a separate content strategy for AI search?
Not a separate one — a stricter one. The same article wins both if it answers a specific question in its opening lines, uses question-shaped headings, and carries facts a model can check. Most sites do not need new pages so much as tighter versions of the pages they have.
Does llms.txt improve my ranking in ChatGPT?
There is no public evidence that it does. It is a sensible, low-cost thing to publish and it may help a model navigate your site, but no assistant has documented it as an input to citation. Do not let it displace structured data or content work.
Can I pay to appear in AI answers?
Not in the organic citations, which is exactly why they are worth earning. Advertising products are appearing inside assistant surfaces, but those are labelled placements and are a separate line item from being cited as a source.
Where to start this week
Pick your five highest-intent pages — the ones a buyer lands on before they contact you. For each one: put a direct answer in the first 40 words, rewrite the H2s as questions, add three FAQs you have genuinely been asked, and ship Article plus Breadcrumb JSON-LD. Then run your prompt panel and write down the result, so next month you have something to compare against.
That is a day of work and it is most of the available upside. Everything after it is iteration — which is the same thing SEO has always been, pointed at a different reader.