How to build an AEO tool

SEO ranks pages. AEO ranks answers.

When ChatGPT answers a question and lists 5 sources, those 5 sources are the new top of page one. The 6th source is invisible. So the question for AEO is not "where do I rank" but "am I in the citation list, and if not, what URLs are taking my slot?"

This post walks through building that tool. Pull citations from every major AI engine, normalize the URLs, build a gap list, run outreach against it, measure the lift. The whole thing is about 150 lines of code.

What an AEO tool actually does

Three jobs, in order of importance.

Citation extraction. For each priority query and each AI engine, pull the list of URLs the engine cited as sources. Normalize them: strip tracker params, resolve redirects, drop fragments.

Citation gap analysis. Find URLs that show up across multiple queries (the "authoritative" sources) and check whether they mention you. URLs that are cited often AND do not mention you are your outreach list.

Citation rate over time. Out of N queries, how often were you cited? Plot that per surface. After outreach, watch the rate climb.

In SEO you optimize the page to win the rank. In AEO you optimize the answer to win the citation. The page is what gets indexed; the answer is what gets used.The AEO mindset shift

Why AEO is harder than it looks

Three operational problems most teams underestimate.

Citation extraction is per-engine work. ChatGPT exposes citations differently from Perplexity, which exposes them differently from Google AI Overviews. Bing Copilot has no public API at all. Wiring up all six takes about a month per engine if you do it right.

URL canonicalization is gnarly. The same article gets cited asexample.com/post, example.com/post?utm=source,example.com/post#section, and www.example.com/post. Treating those as one URL is non-trivial. Maintaining the canonicalization rules as engines change citation formats is ongoing.

API ≠ UI. The citation list returned by ChatGPT's API is often different from what users see in the UI. Per our 1,000-prompt teardown, the API and UI diverge on 96% of ChatGPT queries. Measuring AEO on API-only data is measuring the wrong thing.

The build

import config from "./queries.json" assert { type: "json" };

async function extractCitations(query) {
  const res = await fetch("https://api.mentionsapi.com/v1/check", {
    method: "POST",
    headers: {
      Authorization: `Bearer ${process.env.MENTIONSAPI_KEY}`,
      "Content-Type": "application/json",
    },
    body: JSON.stringify({ mode: "all_live", query }),
  });
  const data = await res.json();
  const out = {};
  for (const [surface, payload] of Object.entries(data.providers)) {
    out[surface] = (payload.citations || []).map(c => ({
      url: c.url,
      domain: new URL(c.url).hostname.replace(/^www\./, ""),
      position: c.position,
    }));
  }
  return out;
}

const allCitations = {};
for (const q of config.queries) {
  allCitations[q] = await extractCitations(q);
}
console.log(JSON.stringify(allCitations, null, 2));

const MY_DOMAIN = "mentionsapi.com";

function citationRate(allCitations, myDomain) {
  const SURFACES = ["chatgpt", "claude", "gemini", "perplexity", "ai_overview", "bing_copilot"];
  const perSurface = {};
  let totalCells = 0, totalHits = 0;

  for (const surface of SURFACES) {
    let hits = 0, total = 0;
    for (const query of Object.keys(allCitations)) {
      const cites = allCitations[query][surface];
      if (cites === undefined) continue;
      total++;
      if (cites.some(c => c.domain === myDomain)) hits++;
    }
    perSurface[surface] = total > 0 ? hits / total : 0;
    totalCells += total;
    totalHits += hits;
  }

  return {
    overall: totalCells > 0 ? totalHits / totalCells : 0,
    perSurface,
  };
}

console.log(citationRate(allCitations, MY_DOMAIN));

function findCitationGaps(allCitations, myDomain) {
  const urlCounts = new Map();
  for (const cites of Object.values(allCitations)) {
    for (const surface of Object.values(cites)) {
      for (const c of surface) urlCounts.set(c.url, (urlCounts.get(c.url) || 0) + 1);
    }
  }
  const popular = [...urlCounts.entries()]
    .filter(([url, count]) => count >= 3 && !url.includes(myDomain))
    .sort((a, b) => b[1] - a[1])
    .slice(0, 50);
  return popular.map(([url, count]) => ({ url, citation_count: count }));
}

const gaps = findCitationGaps(allCitations, "mentionsapi.com");
console.table(gaps);

What this costs

30 queries × $0.50 × 4 weeks = $60 worst case. With 60% cache hits, real cost is $50 a month or less. Add 20 more queries for sample-size reasons and you are still under $100.

The mistake everyone makes

They measure citation rate on day one and then never measure again.

AEO is a longitudinal game. The score you compute today is meaningless in isolation. The score you computed today minus the score you computed 30 days ago is the actual signal. If outreach is working, the line goes up. If it is not, the line is flat and you should change tactics.

Frequently asked questions

Ship it next month

Week one: build the citation extraction pipeline and compute your baseline rate per surface. Week two: build the gap list and pick the top 10 URLs to pitch. Week three to ten: run outreach. Week eleven: re-measure. The lift is the answer to whether your AEO program is working.

Most teams skip the measurement step. Do not be most teams.