AI answers name a few businesses and cite a few pages. The Ranking Factory asks ChatGPT, Gemini, Perplexity, Claude, Grok and DeepSeek the questions your buyers ask, records who they name and which pages they cite, and reads Google’s AI Overviews as a separate surface. Where you are missing, it writes the evidence the question needs — only from your own pages, videos and statements — publishes it on your own site first, and asks again.
Generative Engine Optimization for agencies, consultants and in-house teams. Every number on this site and in the product comes with its sample size and its date, and nothing promises a citation: a measurement, the work, and the proof.
14-day free trial on Growth, then $129 a month. Cancel anytime; a lapsed account goes read-only, and is deleted only after twelve months and thirty days’ notice.
Share of answers that cited a page of your site.
Each point is a score run. The line breaks only where an engine was added or dropped, or the whole question set was replaced; a small tick under the dates marks a model update.
Share of answers that cited a page of your site.
Each point is a score run. The line breaks only where an engine was added or dropped, or the whole question set was replaced; a small tick under the dates marks a model update.
Four of the screens you will use most, captured from a live account. Where a picture is not yet in place you will see a marked box naming the screen — we would rather show a gap than a mock-up.
Measured on ourselves
Read on 6 October 2026 through the same agent read tools a customer’s AI can use on their account, and our site’s own public list of pages. The cards run in the order a citation needs — an engine has to fetch a page before it can cite it — and end with what we have published. The fetched card counts our pages one by one; the cited cards count AI answers and citations that point to therankingfactory.com as a whole, not to the pages on the fetched card. Two caveats travel with them: our own account is the least representative one on the platform — new capabilities are switched on for it first — and every figure says what it was counted over, with the sample beside every rate. Nothing here is a projection.
Fetched
100%
of the 30 live pages our Evidence Engine added to therankingfactory.com between 7 September and 5 October 2026 have been fetched by AI or search crawlers, and ClaudeBot fetched every one. The six guides we published on 5 October were fetched by Bingbot and ClaudeBot the same day.
Source: Agent tool get_published_evidence, project therankingfactory.com, newest 200 ledger items (published 7 Sep – 6 Oct 2026), read 6 Oct 2026; fetches from our Cloudflare zone log, by the names the crawlers give, dated by day
Cited · answers, latest run
83%
of Gemini’s live-search answers cited therankingfactory.com on 6 October 2026 (15 of 18, Gemini 3.8 Flash with Google Search); 15 of those 18 questions carried our name in some form. Perplexity Sonar on Perplexity’s own search and GPT-5.6 Luna on OpenRouter’s web search (OpenAI’s own where the serving host has it, Exa otherwise; this run did not record which), both through OpenRouter, cited us in 56% each (14 of 25; 15 of those 25 questions carried our name in some form). Of the run’s 188 citations of our site, 163 were sources the engines listed and 25 our address in the answer.
Source: Agent tools get_visibility_summary, get_topic_visibility and list_cited_sources, project therankingfactory.com, run of 6 Oct 2026 (25 questions, 13 subjects; OpenRouter models perplexity/sonar and openai/gpt-5.6-luna), read 6 Oct 2026
Cited · answers, like for like
32% → 74%
of Gemini’s live-search answers cited therankingfactory.com: 6 of 19 on 15 August 2026, 14 of 19 on 20 August. Both days used the same model (Gemini 3.7 Flash with Google Search), the same 25-question set and the same counting rule: our address or a link to us in the answer. 16 of the 19 questions carried our name.
Source: Agent tool get_visibility_summary, project therankingfactory.com, runs of 15 and 20 Aug 2026, read 6 Oct 2026
Cited · where the citations go
#1 of 588
sites cited in AI answers to our measured questions: therankingfactory.com, with 971 citations across our 12 latest visibility runs (15 August – 6 October 2026, 27 subjects, on Gemini, OpenAI’s GPT models and Perplexity). 711 of them were sources the engines listed, and the next most-cited site has 297. Most of the questions carried our name.
Source: Agent tool list_cited_sources, project therankingfactory.com, 12 runs, source row for therankingfactory.com, read 6 Oct 2026
Published
50 pages
added to our own site, therankingfactory.com, since 1 September 2026, at least 30 of them by our Evidence Engine (72 pages in all). Evidence pieces support those pages, carrying our own facts in Google Docs, Sheets and Slides and on our YouTube channel: 14 of our videos have been cited in AI answers since 19 September, and in our visibility runs, citations of YouTube sat mostly in answers that already named us.
Source: Public endpoint /api/blog/published (72 pages); agent tools get_published_evidence (newest 200 ledger items, published 7 Sep – 6 Oct 2026) and list_cited_sources (12 runs, the youtube.com row), project therankingfactory.com, read 6 Oct 2026
How a citation is counted, and what it does not claim: /help/how-we-count-a-citation. The measurement method in full: /how-we-measure.
The research
An AI engine that searches the web answers in steps: it finds candidate pages, ranks them, reads a handful and cites some of what it read. A page that drops out at any step cannot be cited in that answer. Researchers at Yonsei University and Konkuk University measured those steps for pages already collected in SAGEO Arena, a benchmark of 2,700 search queries over 171,003 web documents (KDD 2026), on their own search setup rather than Google’s or OpenAI’s. In their benchmark:
−9%
in retrieval when the GEO methods tested rewrote only a page’s body text, on average, and −36% for one of them, AutoGEO.
+22%
in retrieval when only the title, meta description, headings and structured data (JSON-LD) were rewritten.
50–77%
of the target pages that reached the model writing the answer were cited, before any rewriting: 50% by GPT-5-mini, 77% by Claude Sonnet.
Their conclusion: structure and body text do different jobs, and a page needs both — informative body text remains a key factor when pages are ranked and the answer is written.
Source: Kim, Jeong, Kim, Lee and Lee, “SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization”, KDD 2026 (retrieval figures from its Table 2, citation figures from Table 5): arxiv.org/abs/2602.12187.
A page an engine never fetched reaches none of those steps, which is why our site check puts anything that stops a fetch at the top of its list. What we check:
With your Cloudflare zone or our WordPress plugin connected, each page we publish on your site is then placed at one stage — not fetched, refused, crawled, fetched for someone’s question, or cited — and each stage short of cited names its next step. Beside it: whether Google has the page indexed, read through your Search Console, and whether an engine’s search looked at it without citing it, which only some engines report. Where nothing can say, it reads unknown, not no.
Pages first
Anyone can now produce millions of AI-written documents at almost no cost. If that were the way into AI answers, AI slop would become the norm, and Google’s spam policies already have a name for generating many pages to manipulate rankings: scaled content abuse. We are not here to add to it. We count what we publish in subjects, not items, because a second piece saying the same thing in the same place is not more coverage.
Your own site comes first where one is connected: work that makes a good page is prioritised for it. Across our 12 latest visibility runs, our own site drew far more citations than our Google Docs, Sheets and videos put together.
Google Docs, Sheets, Slides and Blogger, in your own Google account, carry the same facts as your pages, where Google’s engines read them. They are your own material: each visibility run’s source count files a cited Doc, Sheet or deck, or a post on a blogspot.com blog, as published for you, never as independent proof about you.
How much, at most: the Evidence Engine’s own limits
Enforced in code and covered by tests, not stated as policy. These are the lines that make the numbers worth reading.
In plain terms: people used to scan a list of links. Now they read one answer. Here is what that means.
The old way
Get a page into the list of ten blue links, and hope people click yours. Still necessary — the fundamentals have not changed — but no longer the whole job.
What changed
ChatGPT, Gemini, Perplexity and Google’s AI Overviews name and cite a few businesses in a single answer. The engines that search the web as they answer read pages published this week; the ones that answer from training cannot, and nothing you publish moves them.
The new way
Know what the engines say when asked, have the evidence they look for published where they read, and be able to prove which of it they fetched and cited. That is what The Ranking Factory does.
Most tools tell you what is wrong. This one measures honestly, does the work, and shows you the receipts.
Six steps, one loop that keeps running for you.
We crawl your site and propose what you do — the kind of business, the places you serve, your topics — each proposal naming the page it was found on. You correct it. The first measurement runs after that, so the baseline is yours, not our guess.
4 of 6 engines per run, your choice, are asked the questions your buyers ask, on your keys. Named, cited and recommended are counted separately, each with its 95% range and sample size; live search and trained recall are kept apart. Google’s AI Overviews are read weekly as their own surface.
Topics where the engines name competitors and not you, evidence the question needs that nothing of yours provides, and every site the answers cited instead — labelled by kind, with links to the exact pages.
Two or three of your own AI providers decide what kind of page the question needs and what it must contain. The writer is handed that contract, your own pages and videos on the subject, and what the strongest outside sources actually say. Invented figures are cut.
Your own site first — WordPress, Shopify, Wix, Ghost or any webhook, signed and idempotent — with the same facts in your Google Docs, Sheets, Slides and Blogger, a video on your channel if you switch videos on, and announcements on LinkedIn or Business Profile only if you opt in. New pages carry provenance, and are announced through IndexNow once your site serves its own key file.
Later runs ask the same questions again and record who cited what. Your Cloudflare zone or our WordPress plugin says which crawlers fetched which page. When a published page drifts from how you describe your product, Re-align rewrites it and waits for your approval; when a late part arrives, the page is updated in place — or, on a page an AI engine is citing, the update waits for your approval.
Agent access
The account answers an AI agent directly over MCP — read and write. Mint a scoped key in Settings and point Claude Code, Claude Desktop, the Claude API or your own tooling at it, or have Claude connect by signing in instead — OAuth with PKCE, no token to copy. Either way you choose at consent whether it reads every project or one; a connection given one project is matched to it exactly and reads nothing else.
Two things travel with every answer on purpose: the sample size beside every rate, so an agent quoting “45% cited” has to quote what it was measured over; and which project the answer covers, because an account-wide total is not a per-client number. No agent can publish, approve or change anything through it unless you tick “Let it make changes” when you make the key or the connection — off by default, audited, and capped at 30 writes an hour once it is on. Both ways in are on every plan, Starter included, and in the trial.
Eleven read tools, four scopes granted separately
Ask for a project that does not exist and you get an error naming the ones that do — never a silent empty list.
One project per client domain, each with its own measurements, publishing destination, topic list and “needs you” list. Several branches or divisions go in together and each can become its own campaign. White-label reports and client-ready activity reports from Agency; one weekly email per client site, on by default, says what was cited, published, crawled and indexed, and a client can get its own proof-only edition. Connect Claude by signing in and scope it to one client.
See agency featuresStart the 14-day Growth trial. The first thing it does is check your site, which needs nothing but a web address; the second is ask the engines about you and show you the answers with their sample sizes.