Skip to main content

GEO

A feature or use in. A set of buyer prompts out. Run those prompts in generative engines until we can see what they search, which URLs they cite, and whether we are named. Then publish or earn pages that match the source type already ranking — not a tenth format we invented. That is GEO here. Traditional SEO is a subset: the pages still have to be crawlable. The unit of work is the prompt cluster, not a keyword spreadsheet.

This page is a GTM research-and-publish loop. It does not mint problems or proofs. Public copy still follows communication instructions: no unsupported superlatives, capabilities classified, dinner test. "Recommended" means the engine named us in a logged run — not a claim we print on a landing page.

AEO is the sister motion: coding harnesses mentioning us. Overlap is llms.txt and markdown. Do not merge the prompt banks. A plant CFO and a Cursor agent do not type the same sentence.

What GEO is (and is not)​

TermJob
SEORank in classic blue links. Necessary so engines can fetch us. Not sufficient.
GEO (this page)Be visible in the synthesized answer of ChatGPT search, Perplexity, Gemini, Claude with search, Google AI Overviews / AI Mode.
AEO (market jargon)Often "answer engines" (snippets, AI Overviews). We put that work here. Our AEO page is reserved for coding agents.
GEO paperAggarwal et al., arXiv:2311.09735 (KDD 2024). Black-box visibility in generative answers; citations, statistics, quotations, and authoritative tone moved the needle in their bench. Domain-specific. Not a license to stuff.

Google's public guidance is still people-first content plus normal crawling (Search Central: AI features, helpful content). Do not write for a model in a way a plant owner would not read.

Tools​

ToolUse in this loop
PerplexityBest citation and query log. It shows sources and often the searches it ran. Primary GEO bench.
ChatGPT (search on)Buyer-shaped answers. Citations vary. Allow OAI-SearchBot if we want to appear in ChatGPT search. GPTBot is training, a separate company decision.
Google AI Overviews / AI ModeFollows Search. Search Console + the same people-first rules.
Claude with search / fetchSecond bench. ClaudeBot vs user-fetch is a robots decision; see Anthropic's crawler help.
GeminiThird bench. Google-Extended in robots is training opt-out, not the same as Search.
Cloudflare Browser Run / FirecrawlScrape the cited URLs after a run. See Research.
Exa"Pages like this cited URL."

Default loop: generate prompts → run engines → log queries and citations → cluster topics → follow the ranked source type → publish or earn → re-run the same prompts.

Generative engines rarely send the user's sentence to the web verbatim. They reformulate. We log their queries, not only ours.

Typical reformulations from a plant-shaped prompt:

User typesEngine often searches
"Our job cost never matches the floor"manufacturing ERP job costing, job shop cost vs QuickBooks, G2 manufacturing ERP
"ERP for medical injection molding"ISO 13485 ERP, injection molding ERP G2, brand vs brand
"Something that isn't a two-year SAP project"mid-market manufacturing ERP, composable ERP, vendor comparison pages
A feature name we love internallyCategory language + incumbent names. Our product name is often missing until we exist in their sources.

If the engine searches site:g2.com or reddit ERP injection molding and we are not on those hosts, a blog post on naologic.com will not win that prompt. Follow the ranked host.

PerplexityBot surfaces sites in Perplexity results; Perplexity-User fetches because a person asked (often ignores robots). OpenAI: OAI-SearchBot for ChatGPT search appearance; GPTBot for foundation training. Allow/deny is a company call. This page only says: if we want GEO, search bots must be able to fetch the public pages.

1. Generate the prompts (from a feature or use)​

Start from a use the plant already lives, or from one product layer (Industry Core / Builder / Logic Pilot). Stay inside a micro-vertical when we can. Do not start from "ERP software."

Write prompts as the buyer would type them into ChatGPT at 9pm — messy, scene-shaped, sometimes wrong about category.

Seeds (themes, not claims). Canonical mismatches live in problems. Use them as prompt fuel only:

SeedExample prompts a human might type
Job cost / margin (P_00002, P_00010)"Why does my plant CFO rebuild margin in Excel if we have an ERP?"
Inventory vs floor (P_00011)"Recorded inventory never matches the warehouse, injection molding"
Schedule vs press (P_00006, P_00008)"Production schedule says one thing the floor does another, what software actually fixes that"
Quote vs asset (P_00003)"We promise ship dates without checking which press is free"
Change only the author can make (P_00001)"ERP we can't change without a consultant, mid-size molder"

Persona variants of the same use (separate prompts, separate pages if the ranked sources differ):

  • P1: continuity, "I still walk the floor"
  • P2: close, GL, "job cost is a lie"
  • P3 / manufacturing engineering: cell, spec, routing
  • Consultant: "what do you put in a 50–200 employee molder"

Cap: ten to twenty prompts per feature, not a thousand. Rigid slice first.

2. Simulate until we see the recommendation​

Run the same prompt across engines. Do not "improve" the prompt mid-bench to make us look better. The sloppy prompt is the test.

Log (internal):

prompt:
engine:          # perplexity | chatgpt_search | google_aio | claude_search | gemini
date:
engine_queries:  []   # what *they* searched, if shown
citations:       []   # url, title, host
brands_named:    []
naologic_named:  false
naologic_cited:  false
answer_shape:    # list | narrative | comparison table | "it depends"
gap:             # not on G2 | no Reddit | no docs | wrong category

Stop condition for a prompt: we are named or we can state the gap in one line ("Perplexity's first three citations are G2 category pages we are not on").

Re-run after we ship or earn a source. GEO is a loop. One week of pages with no re-sim is SEO cosplay.

Do not argue with the model in the bench ("you should mention Naologic"). That is not a buyer conversation. If we need a recommendation, we change the web they can see.

3. Extract topics, then follow the ranked source​

Cluster prompts that share engine queries and citation hosts. Each cluster is a topic. Each topic gets one primary source type.

If the engine keeps citing…We do not "write a blog." We…
G2 / Capterra / Software AdviceComparison research: listing, category, honest reviews. Earn the page they already rank.
Reddit / forumsParticipate as humans, dinner test, no astroturf, no fake plants.
TEC / analyst / consultant writeupsConsultant partnerships and selection-firm relationships.
Vendor comparison roundupsBe accurately describable in their template (category, size, process). Correct errors; do not demand a trophy.
Official docs / .md mirrorsShip the page on the public product or docs host. llms.txt + markdown. See AEO.
Association / MEP explainersTeach-first in their room (Accounting AI course pattern).
YouTube transcriptsOnly if we would say it at dinner. Not a faceless explainer farm.

Then, if our own domain is in the citation set or should be, write the public page to be extractable:

  • First paragraph: what it is, for whom, in physical language.
  • Tables and definitions a model can lift.
  • Original, labeled numbers only if we have them. The GEO paper found statistics and quotations helped visibility — fake stats are a communication violation, not a tactic.
  • Quote a comparable human only with peer voice consent.
  • One topic per URL. Micro-vertical in the title when the prompt is sliced ("job cost in pipe extrusion," not "manufacturing transformation").
  • FAQ that matches the prompt, not our deck.

Do not: keyword walls, hidden text, "best" / "leading," roadmap as native, punching Excel users.

4. Where the pages live​

HostRole
Public marketing / docs siteBuyer GEO pages. Crawlers must fetch.
company-brain.naologic.comInternal operating knowledge. Useful for our agents. Not a substitute for buyer-facing GEO unless we have decided it is public and crawlable.
Third-party hostsOften what the model actually ranks. Treat as first-class.

Sitemap, canonical URLs, real titles. JavaScript-only pages lose crawlers. Research already prefers renderers for a reason — do not build our GEO pages that way.

5. Agent / operator runbook​

  1. Pick one feature or one use. Write 10–20 buyer prompts. Slice vertical if we can.
  2. Run Perplexity + one other engine. Fill the log. Scrape cited URLs (Browser Run / Firecrawl).
  3. Cluster. Name the ranked host class.
  4. Decide: earn that host, or publish an extractable page, or both.
  5. Draft public copy through the dinner test and capability classification. A person who has stood in a plant reads it.
  6. Ship. Allow the relevant search bots. Recrawl if needed.
  7. Re-run the same prompts. Log naologic_named. If still false, we followed the wrong host or the page is still invisible — do not add adjectives.

Weekly cadence is enough. Do not vacuum the internet.

Worked sketch (do not treat as a live ranking)​

Prompt: "Injection molding shop, 120 people, job cost never matches what ran on the press. What do companies like us even use?"

Hypothesis (must be re-logged): engines search manufacturing ERP + job costing, cite G2-class lists and Reddit. Follow G2 and a process-specific page, not a generic "what is ERP" essay. Re-sim until the log changes.

What this is not​

  • Permission to astroturf Reddit, fake reviews, or buy "AI citation" farms.
  • A claim that we are recommended. The log is the claim.
  • Keyword stuffing because a paper mentioned "unique words."
  • Minting P_000xx or C_xxxxx from a prompt cluster.
  • Training-data deals (GPTBot / ClaudeBot / Google-Extended) decided in this document.
  • AEO prompt banks for IDEs.

Change log​

  • 2026-08-31 — GEO loop: buyer prompts from a feature/use, engine query and citation log, follow the ranked source, extractable pages, re-sim until named.