AI Assistants Quote Pages That Answer in the First 2 Sentences. I Restructured 3 Pages to Test It.

Ask Perplexity what a website costs in India. It gives you a range and cites 3 or 4 pages. Now open those pages and read them. Every single one answers the question inside the first two sentences, with an actual number sitting right there.

I kept seeing this pattern once I started looking for it. In June I audited webight.com properly (that audit found a 4,190ms render delay on my own hero headline, separate story) and afterwards I went down a different rabbit hole: which pages do ChatGPT, Perplexity and Gemini actually quote when they cite sources? The same shape showed up again and again. A paragraph that stands completely on its own, answers one question directly, and carries a specific fact. Usually a number.

So I restructured 3 pages on my own site to match that shape, left a control page untouched, and set up logging to watch what happens. This post is the test. The before paragraph, the after paragraph, the setup, the measurements, and the early numbers.

Why a paragraph gets lifted

When an AI assistant answers with citations, it works on chunks. Your page gets split into pieces, each piece gets scored against the user's question, and the highest scoring piece gets quoted or paraphrased into the final answer. The assistant never appreciates your brand story on its way down the page. It grabs the one block of text that resolves the query and moves on.

That changes what a "good first paragraph" means. A paragraph that needs the paragraph above it to make sense scores badly, because on its own it answers nothing. A paragraph that opens with the answer and backs it with a number scores well, because the retrieval system can hand it to the user as-is.

There is a human reason underneath this too. Concrete figures read as more credible than vague claims. Tell someone "websites are affordable" and they shrug. Tell them "a landing page starts at INR 7,999" and they believe you, because a specific number sounds like it came from somewhere real. The models were trained on human preferences, so they inherited the same bias. Sentences with exact figures survive into the final answer. Vague ones get dropped during synthesis.

See, this is the part most site owners miss. They write the first paragraph for a visitor who will scroll. The machines that now decide whether anyone visits at all read the first paragraph as a standalone snippet. Two different readers. The page has to serve the second one first.

The before paragraph

This was the top of my pricing page until last month, lightly paraphrased:

Every business is different, and so is every website. We believe good design starts with understanding your goals, which is why our plans are flexible and built around what you actually need. Get in touch and we will figure out the right fit together.

Read that as a retrieval chunk. What question does it answer? None. There is no price, no timeline, no fact. It exists to warm the visitor up, and warmth scores zero against a query like "how much does a website cost in India". If an assistant pulled this chunk, it could not build a single sentence of an answer from it. So it never pulls it.

I wrote paragraphs like this for two years. They felt professional. They were invisible.

The after paragraph

Here is what sits at the top of that page now:

A landing page at Webight starts at INR 7,999 ($149) and a full website starts at INR 24,999. Payment is 50% upfront and 50% on delivery, with a 7-day money-back window after kickoff.

Two sentences. Five numbers. It answers "how much does a website cost" without needing anything around it. Paste it alone into a chat and it still works. That is the whole test for this shape: the paragraph has to survive being ripped out of the page.

The brand story did not get deleted. It moved down. Humans who keep scrolling still get the warm version. The chunk at the top does the machine-facing work.

The 3-page test

Three pages, each rewritten to answer one target question in its first two sentences, each with a number.

Page 1: /pricing. Target question: "how much does a website cost in India". First paragraph rewritten as above.

Page 2: /services, the automation section. Target question: "do WhatsApp booking bots actually work for hotels". New opening: "A WhatsApp and LINE booking bot we built for a boutique hotel produced 600+ direct bookings with zero ad spend. The bot answers guests inside the chat and completes the booking there, so the hotel pays no OTA commission on those stays."

Page 3: a blog post on slow hero sections. Target question: "what causes a long LCP element render delay". New opening: "The most common cause of a huge LCP render delay on a Next.js site is an entrance animation waiting for hydration. On my own homepage, a framer-motion opacity animation held the hero headline for 4,190ms; replacing it with a plain CSS keyframe animation (translate only, no opacity, no JavaScript) dropped the delay to 382ms."

The control is my about page plus everything else on the site, all left exactly as they were. If the 3 restructured pages get crawled harder and cited while the control sits flat, the shape did it.

One honest caveat before the numbers. Three pages on one small site is a directional test. I am watching for a clear difference between treated and untreated pages, and I am not going to claim laboratory proof from n=3.

How I measure it

Two things: crawler behaviour and actual citations.

For crawlers, a Next.js middleware logs every AI bot hit into Supabase. Complete code, drop it in and deploy:

// middleware.ts
import { NextResponse } from "next/server";
import type { NextRequest, NextFetchEvent } from "next/server";

const AI_BOTS = [
  "GPTBot",
  "OAI-SearchBot",
  "ChatGPT-User",
  "ClaudeBot",
  "Claude-User",
  "PerplexityBot",
  "Perplexity-User",
  "CCBot",
  "Bytespider",
];

export function middleware(req: NextRequest, event: NextFetchEvent) {
  const ua = req.headers.get("user-agent") ?? "";
  const bot = AI_BOTS.find((b) => ua.includes(b));

  if (bot) {
    event.waitUntil(
      fetch(`${process.env.SUPABASE_URL}/rest/v1/ai_crawler_hits`, {
        method: "POST",
        headers: {
          apikey: process.env.SUPABASE_SERVICE_ROLE_KEY!,
          Authorization: `Bearer ${process.env.SUPABASE_SERVICE_ROLE_KEY!}`,
          "Content-Type": "application/json",
          Prefer: "return=minimal",
        },
        body: JSON.stringify({
          bot,
          path: req.nextUrl.pathname,
          ua,
        }),
      })
    );
  }

  return NextResponse.next();
}

export const config = {
  matcher: ["/((?!_next|api|.*\\..*).*)"],
};

And the table:

create table ai_crawler_hits (
  id bigint generated always as identity primary key,
  bot text not null,
  path text not null,
  ua text,
  hit_at timestamptz default now()
);

For citations, I keep a fixed list of 9 prompts, 3 per page, phrased the way a real person would ask. Every Sunday I run all 9 in fresh sessions on ChatGPT with search, Perplexity, and Gemini, and record four columns in a sheet: date, assistant, was webight.com cited, and the exact sentence that got quoted or paraphrased. That last column matters most. It tells you which paragraph shape gets lifted, which is the entire question this test exists to answer.

That's it. No tool subscription, no dashboard. A middleware, a table, and a boring spreadsheet.

What happened so far

37 days in, early and small-scale:

The 3 restructured pages have taken 47 AI crawler hits combined against 9 on the about page.

Perplexity cited the pricing page in week 3 and quoted the first sentence nearly word for word, INR figure included.

Gemini paraphrased the 4,190ms number from the blog post once in week 5. ChatGPT has cited nothing yet.

No traffic flood. No lead I can attribute to it yet. I want to be straight about that, because the quoting behaviour is what moved, and the quoting behaviour is what I set out to move. Whether citations turn into clients is the next 6 months of the spreadsheet.

What I can already tell you with confidence: the paragraphs that got lifted were the ones that survive standing alone. Both quotes came from the exact two sentences I rewrote. Nothing from page bodies, nothing from the warm-up copy lower down, nothing from the control.

So here is your move. Open your most important page and paste its first paragraph, alone, into a chat with a friend. No context, no link. If they cannot tell what question it answers, an AI assistant cannot either, and that paragraph will keep losing citations to whoever put their number in sentence one. Rewrite it this week, log your crawlers, and check back in a month.

If you run this on your own site, DM me what your logs show. I want more than n=3.