What makes ChatGPT, Perplexity, Copilot, and Claude recommend a site

ChatGPT, Perplexity, Copilot, and Claude recommend a site when they can fetch it, lift a consistent entity fact from the HTML, see other domains repeat that fact, and treat the claim as current. That is a recipe, not a leaked ranking formula. Vendors publish crawler tokens and product behavior. They do not publish a score you can buy.

This is the recommendation recipe under the Answer Engine Optimization playbook. Which engine to prioritize is a surface map: ChatGPT vs Perplexity vs AI Overviews. Why ChatGPT names everyone else is a diagnostic: why ChatGPT recommends competitors. This spoke is what to build so a recommendation is even possible.

I have been SEO certified since 2021. The AEO version of that work is still eligibility plus extractable truth. I will not invent a secret ranking factor for your deck.

The short answer

• Four ingredients: crawl access, extractable entity facts, third-party corroboration, recency. Miss one and the recommendation usually fails before “content strategy” starts.
• Search and training are different jobs. GPTBot and ClaudeBot train. Citations ride search and user-fetch twins: OAI-SearchBot, PerplexityBot, Bingbot, Claude-SearchBot, Claude-User.
• ChatGPT mixes training memory with selective search. Perplexity is retrieval-first. Copilot grounds public-web answers in Bing. Claude splits training from search and live fetch.
• Shared work transfers. Retrieval neighborhoods do not. Ahrefs measured almost no shared top sources across ChatGPT, Perplexity, and Google AI Overviews in a June 2025 cut.
• Measure a frozen prompt panel on all four. One lucky ChatGPT run is not a program.

What actually makes them recommend a site?

A recommendation is a model (or its search layer) deciding it is safe to name you as an option. Safety here is boring: the system can retrieve you, compress you into a sentence, and defend that sentence against contradictory pages. It is not a trophy, a submission form, or a paid inclusion button for ordinary sites.

Vendors do not publish the weights. Treat the table as an operator model.

Step	What happens	What you control	What you do not control
1. Fetch	A bot or a user-triggered agent can read the URL	robots.txt, WAF, HTTP 200, no login wall	Crawl budget, fan-out queries
2. Extract	A passage states who you are, what you sell, for whom	First-screen HTML, tables, matching schema	How aggressively the model compresses
3. Corroborate	Other domains repeat the same entity facts	Directories, reviews, roundups, press, docs	Whether that domain is in this engine’s pool
4. Date	The claim still matches the live offer	Visible dates, IndexNow pings, killing stale URLs	Recrawl SLA (none is published as a citation clock)
5. Attribute	The product shows a source, a name, or both	Quoteable sentences with your name in them	Citation vs mention vs ghost citation

If step 1 fails, nothing else matters. If step 1 works and steps 2–4 conflict, the system often names a competitor with a cleaner trail — that diagnostic lives on the competitors spoke, not here.

Recommendation prompts (“who should I hire,” “best X for Y,” “alternatives to Z”) are stricter than definition prompts. A definition can cite a glossary. A recommendation has to justify putting you on a shortlist. That is why corroboration and recency punch above word count.

• You can say in one sentence what a “win” is: named, cited, or both
• The money URL returns 200 to a fetch that does not execute your marketing JS
• About, offer, and schema agree on the same legal name and category noun
• At least a few independent URLs repeat those facts
• Pricing and service-area claims on the live page match the last 90 days

A pretty site with a blocked search bot is a brochure the engine never opened.

The four ingredients, in order

Do them in this order. Reversing the order is how teams ship a 40-post calendar nobody can fetch.

Order	Ingredient	Pass test	Fail look
1	Crawl access	Citation-useful bots get 200 on About, offer, FAQ	“Block AI” robots.txt, WAF challenge, JS-only text
2	Extractable entity facts	First 60–80 words state who / what / for whom; schema matches	Hero slogan, then three screens of memoir
3	Third-party corroboration	Independent pages repeat name, category, geo	Only your domain and your LinkedIn
4	Recency	Live price, offer, and hours match the HTML a bot will copy	Old packages still ranking in Bing or in training residue

Numbered procedure:

1. Allow the citation tokens. Leave training tokens as a written policy call.
2. Put one entity packet on About + homepage + offer + JSON-LD. Same strings.
3. Get those strings repeated off-site. Directories and honest reviews before a thought-leadership sprint.
4. Date the claims that change. Kill or 301 the URLs that still sell the old offer.
5. Re-run the same prompts. Do not rotate the questions to chase a screenshot.

Shared baseline vs engine-specific bets:

Work	Transfers across all four	Engine-specific
Entity packet	Yes	No
Answer-first money URL	Yes	No
robots.txt / WAF for that vendor’s search bot	Per token	Yes — different tokens
Bing index + IndexNow	ChatGPT Search (partnered) and Copilot	Copilot especially
Community threads (Reddit, forums)	Corroboration layer	Heavier in some Perplexity and ChatGPT studies
Prompt panel	Method transfers	Four logs, not one blended “AI score”

Cap engine-specific experiments after the four ingredients pass. A Perplexity-only blog cluster on a site PerplexityBot cannot fetch is theater.

Search vs training: the split operators keep collapsing

This is the mechanical difference that wrecks otherwise competent teams. Training crawlers collect pages that may enter foundation-model datasets. Search crawlers and user-fetch agents decide whether a live answer can retrieve and cite you. OpenAI is explicit that the two robots.txt tags are independent: you can allow OAI-SearchBot for ChatGPT search results and disallow GPTBot so crawled content should not be used to train generative AI foundation models (OpenAI crawler docs). Anthropic publishes the same split across ClaudeBot (training), Claude-SearchBot (search quality), and Claude-User (user-initiated fetch) (Anthropic Help Center).

Perplexity’s published position is simpler: PerplexityBot surfaces and links sites in search results and is not used to crawl content for AI foundation models (Perplexity crawlers). Copilot does not ship a “CopilotBot.” Public-web grounding goes through Bing (Microsoft Support: Copilot web search).

Product	Training / memory path	Live recommendation path	Collapse this and you…
ChatGPT	GPTBot; answers that never trigger search	OAI-SearchBot for search answers; ChatGPT-User for some user fetches	Block search while celebrating a training opt-out
Perplexity	Not what PerplexityBot is for, per vendor	PerplexityBot index + Perplexity-User live fetch	Treat Perplexity as “another GPTBot”
Copilot	Model weights; M365 work data if Work IQ is on	Short Bing query from the prompt	Optimize Google and skip Bing
Claude	ClaudeBot	Claude-SearchBot + Claude-User	Disallow the wrong token and call it privacy

Hedge: if you allow both of OpenAI’s tags, OpenAI says it may use results from just one crawl for both use cases to avoid duplicative crawling. That is a crawl-efficiency note, not a reason to treat the tags as one switch.

Training residue still matters. An older Claude or ChatGPT answer can name a product you killed in 2024 until live search overrides it — or until you never trigger search and the stale sentence wins. Recency (ingredient 4) is how you fight residue. Blocking the training bot does not erase what already landed.

Panel condition	You are looking at	Recipe move
Search / web is on and sources appear	Live retrieval	Ingredients 1–4 on the cited neighborhood
Search / web is on and zero sources	Memory or a failed fetch	Confirm the product actually searched; then crawl
Search / web is off	Training / product default	Turn it on for the panel or stop claiming a “search citation”
Two engines disagree on the same prompt	Different pools, same recipe	Do not average them into one KPI

Copilot adds a fifth fork: web search can be off at the tenant. Microsoft documents that org policy can serve answers “without the benefit of current web information.” Your public pages cannot override that toggle. Log whether web was on.

Crawl access: eligibility, not a ranking secret

Crawl access is a gate. It is not a ranking factor I am pretending to have leaked. OpenAI’s wording is the template: sites opted out of OAI-SearchBot “will not be shown in ChatGPT search answers, though can still appear as navigational links,” and it can take about 24 hours after a robots.txt change for search systems to adjust (OpenAI crawler docs). Perplexity also says robots.txt changes may take up to 24 hours (Perplexity crawlers).

Match the product token, not a pinned version string. Version suffixes change.

Engine	Citation-useful token	Training / other token	User-triggered fetch	robots.txt note
ChatGPT Search	OAI-SearchBot	GPTBot	ChatGPT-User	User fetch: rules may not apply; Search opt-out is OAI-SearchBot
Perplexity	PerplexityBot	Not this bot’s job, per vendor	Perplexity-User	User fetch generally ignores robots.txt; whitelist WAF IPs
Copilot	bingbot	n/a as a Copilot crawler	Bing retrieval	Copilot sends a short query to Bing
Claude	Claude-SearchBot	ClaudeBot	Claude-User	Disabling SearchBot or User “may reduce” visibility, per Anthropic

Bing’s own crawler list names Bingbot as the standard crawler (Bing Webmaster: which crawlers). Microsoft Copilot Chat generates a short Bing query from the prompt and shows a Sources button with that query (Microsoft Support). There is no separate Copilot index you submit to.

WAF and CDN “AI bot” bundles are where eligibility dies silently.

• robots.txt allows OAI-SearchBot, PerplexityBot, bingbot, Claude-SearchBot, Claude-User on public money URLs
• Training tokens (GPTBot, ClaudeBot) are Allow or Disallow on purpose, with a date and an owner
• Cloudflare / AWS managed lists do not dump citation crawlers into a scrape bundle
• Logs or WAF events show 200s, not JS challenges, on About and the offer URL
• You verified IPs against vendor JSON where they publish it — UA strings are spoofable

Published IP lists (re-fetch; do not pin a copied range in a ticket from 2024):

Token	Vendor IP JSON	Use it for
OAI-SearchBot	openai.com/searchbot.json	WAF allow + log authenticity
GPTBot	openai.com/gptbot.json	Training policy only
ChatGPT-User	openai.com/chatgpt-user.json	User fetch; not Search opt-out
PerplexityBot	perplexity.com/perplexitybot.json	Index crawl allowlist
Perplexity-User	perplexity.com/perplexity-user.json	Live question fetch
bingbot	Bing’s Verify Bingbot tooling	Copilot web grounding
Anthropic bots	IP list on their crawler help page	Do not IP-block as the opt-out — they say it may fail

OpenAI: ChatGPT-User is not used to determine whether content appears in Search. Anthropic: blocking by IP may not persist, because it can stop the bot from reading robots.txt. Perplexity: combine User-Agent and published IP ranges in the WAF. Those are vendor notes, not folklore.

A blocked search bot is not a ranking problem. It is a missing fetch.

Extractable entity facts the model can quote

If the crawler gets a 200 and the first screen is a slogan, you failed ingredient 2. Answer engines cite passages they can lift. Google’s AI-features guidance is useful even when this post is not an Overviews tutorial: important content has to exist as text, and structured data should match the visible page (Google Search Central: AI features). schema.org’s Organization type is the shared vocabulary for name, URL, logo, and related identifiers (schema.org/Organization). Google’s Organization guide says homepage markup helps disambiguate the entity; add the properties that apply, and do not contradict the HTML (Google: Organization structured data).

Fact	Put it in HTML as a sentence	Also in JSON-LD	Do not hide it in
Legal + trade name	First screen of About and homepage	name / legalName / alternateName	A footer SVG
Category noun	“We are a {category} for {ICP}”	knowsAbout or a clear description	“Solutions for tomorrow”
Offer / pricing	One current package sentence	Offer markup only if it matches	A PDF sales deck
Geo / service area	Plain sentence	areaServed / address	A map widget with no text
Founding / identity	One date, one city	foundingDate	Three conflicting blog asides

Quote test — 60–80 words, no hero video required:

1. Open the money URL with JS off, or view source.
2. Copy the first 80 words of visible text.
3. Ask: would this paragraph still be true if a model cited only that span?
4. If the span is a metaphor, rewrite it as a fact.
5. Check that Organization JSON-LD uses the same strings.

I have shipped hundreds of production sites. The pages that get named in recommendation answers are boring on purpose: noun, buyer, job, proof, next step. Cinematic heroes convert humans. They do not give a retrieval system a sentence it can defend.

Page	Extractable job	Failure
Homepage	Who, what, for whom, where	Atmosphere without a category
About	Entity packet	Team photos, no legal name
Service / product	Criteria + scope	Feature salad
FAQ / proof	Objection-shaped facts	Marketing FAQ with no numbers
Contact / location	NAP that matches directories	Five addresses, none canonical

Conflicting facts are worse than missing facts. A model that sees two founding years often stays conservative — or invents a third. Align the packet before you write more URLs.

Third-party corroboration the homepage cannot fake

Your homepage is one node. It is not the trail. Recommendation prompts look like “who is a real option in this category.” Independent repeats — directories, reviews with job language, roundups, association lists, trade press, docs on other domains — are how a system checks that you exist outside your own DNS.

This is not a secret “off-site ranking factor.” It is the same corroboration problem as the competitors diagnostic, pointed at construction instead of exclusion. Ahrefs’ June 2025 Brand Radar cut found 86% of top mentioned sources were not shared across ChatGPT, Perplexity, and Google AI Overviews, with only seven websites in the top 50 for all three (Ahrefs). Read that as: corroboration has to land in that engine’s neighborhood, not only on a domain you like.

Corroboration type	What a model can use	What does not count
Directory / profile	Consistent NAP + category	Five profiles, five categories
Review	Specific job-to-be-done language	“Great service!!!”
Roundup / “best of”	You listed with the same category noun	A paid DR package on an unread blog
Press / association	Named, dated, checkable	Your own Medium cross-post
Community thread	A real answer other people already cite	Spam seeded for Perplexity

Semrush’s February 2026 SaaS sample compared Google’s top 10 pages against pages cited by AI products: ChatGPT overlap with that top 10 was 2.1% — the lowest of the platforms they listed (Semrush: AI visibility). Sample: 10 SaaS queries. Treat it as a measured miss, not a universal law. It is still why “we rank on Google” is not a recommendation recipe.

Ahrefs’ 75,000-brand analysis reported branded web mentions correlating more strongly with AI Overview brand visibility than backlink count (Spearman 0.664 vs 0.218) — correlation, not a causal citation API (Ahrefs: AI Overview brand correlation). Mentions on pages engines already retrieve beat a bought Domain Rating report.

• One public name and one category noun on owned pages
• At least a handful of third-party URLs that repeat those strings
• Reviews (if you collect them) mention the job, not only the vibe
• You are not paying for links on domains your prompt panel never cites
• You did not spam Reddit because a screenshot showed Reddit in Perplexity footnotes

Fake reviews and directory stuffing are not corroboration. They are a trust problem you will still be paying for after the mention fades.

Recency: stale offers get skipped or invented

Recency is not “blog more.” It is “the live claim matches the retrieved claim.” Pricing pages, retainers, service areas, and product names go stale. Training residue and old press releases keep the corpse warm. Recommendation answers that quote a dead package either skip you or hallucinate you into the old offer.

Google’s recrawl language is the honest clock: crawling can take several days to several months depending on how often systems decide a page needs a refresh (Google AI features). IndexNow’s FAQ is equally honest: submitting a URL does not guarantee indexing (IndexNow FAQ). OpenAI and Perplexity quote ~24 hours for robots.txt adjustments, not for “you will be recommended tomorrow.”

Claim type	Refresh when	Recency move	Not a recency move
Pricing / packages	The number changes	Ship HTML the same day; ping IndexNow	Bump lastmod with no edit
Service area	You add or drop a city	One sentence on location + directories	A new blog post titled “we’re expanding”
Product name	You rename	301 the old URL; update roundups you can reach	Leave the old URL as a “legacy” ranking play
Evergreen explainer	The answer is wrong	Rewrite the lift-able span	Quarterly rewrite of a winning passage
News / research	The window expires	Date the method; demote or archive	Keep “2023 survey” in the H1 in 2026

Bing’s AI Performance preview tells publishers that accurate, up-to-date content matters for inclusion in AI-generated answers and points at IndexNow for faster discovery (Bing Webmaster: AI Performance). That is a freshness hygiene note. It is not a Copilot citation SLA.

Microsoft’s consumer Copilot transparency note says that when Copilot is grounded in web search, it “centers its response on high-ranking content from the web” and attaches hyperlinked citations (Microsoft Copilot transparency note). “High-ranking” is their language. It is not a published formula that “Bing #3 = Copilot recommendation.” Hedge it.

Recency failure	What the answer does	First ticket
Old price still indexed	Quotes the corpse or skips you	Update HTML + IndexNow + kill mirrors
New offer, old About	Conflicting entity	Align the packet
Training residue only	Names a product you sunset	Live search path + third-party updates
Date theater	dateModified with no change	Real edit or leave it alone

Do not rewrite a passage that already gets quoted. Recency is accuracy, not fidgeting.

How the recipe lands on ChatGPT vs Perplexity vs Copilot vs Claude

Same four ingredients. Four retrieval neighborhoods. Do not run four content calendars.

Engine	When it recommends from the live web	When it recommends from memory	Operator bet after the four ingredients
ChatGPT	Search ran; OAI-SearchBot could index you; partner queries (Bing is named) returned corroborable pages	Search did not run; training residue answers	Allow OAI-SearchBot; Bing eligibility; quoteable criteria page
Perplexity	Almost every factual query retrieves; numbered citations to live URLs	Rare for “who should I hire” if retrieval is on	Allow PerplexityBot; unique HTML facts; WAF allowlist
Copilot	Web search on; Bing returns you in the rewritten query	Work IQ / files, or web search off	Bingbot + Webmaster Tools + IndexNow; spot-check if buyers live in M365
Claude	Web tools ran; SearchBot indexed; User could fetch	ClaudeBot-era residue; no web	Allow Claude-SearchBot and Claude-User; do not confuse with ClaudeBot

ChatGPT Search rewrites the prompt into one or more targeted queries and sends those to partner search providers, including Bing (OpenAI Help: ChatGPT Search). Independent August 2026 capture work argued OpenAI also serves an in-house index that does not match Bing’s top 20 on the same fan-outs (Search Engine Land / RESONEO). OpenAI has not documented that pipeline name. Hedge it. Keep Bing hygiene. Do not treat “ChatGPT = Bing only” as settled doctrine.

Perplexity describes the product as searching the internet in real time and attaching numbered citations (Perplexity Help). Third-party studies often find community URLs (notably Reddit) in the mix; percentages move. Earn those threads. Do not spam them.

Copilot: if your organization turns web search off, you get a model answer without current web insights (Microsoft Support). Your public-site recipe cannot override an admin toggle. Measure the consumer / web-on path your buyers actually use.

Claude: Anthropic’s own disable-language is the hedge — restricting SearchBot or User may reduce visibility and accuracy. That is not a published citation rate.

Do this once	Do not do this four times
Entity packet + five quoteable URLs	Four “AEO programs”
Token table with owners	Four robots.txt religions
One frozen prompt list	Four vanity dashboards
Corroboration on domains the panel already cites	Four Reddit sprints

Priority still follows buyers. The engine map decides where to spend extra. The recipe decides whether you are eligible anywhere.

Worked example: one hire prompt through the four ingredients

Prompt (freeze it): “Who should I hire for {category} in {metro}?”

Walk the recipe. Do not skip to a blog outline.

Ingredient	Check on your site this afternoon	What the four engines do with a miss
Crawl	curl -A the citation UA (or check logs) on / and the service URL	ChatGPT Search never shows you; Perplexity never footnotes you; Copilot never sees a Bing hit; Claude User/Search never fetch
Extract	First 80 words of the service URL name the category, ICP, and metro	Retrieved but not shortlisted — slogan is not a criterion
Corroborate	Three independent URLs use the same category + metro	ChatGPT names directory-backed rivals; Perplexity footnotes a roundup you are not on
Recency	Price, hours, and service area match GBP and the HTML	Copilot quotes last year’s package from Bing; Claude memory names a sunset SKU

Procedure for that one prompt:

1. Run it in all four products with search / web on. Save the source list.
2. Circle every cited domain. Mark owned vs third-party.
3. If you are absent and rivals’ directory pages appear, you have a corroboration gap, not a headline gap.
4. If you are absent and no small sites appear, check fetch first.
5. If you are named with the wrong offer, it is recency or a conflicting packet — not “we need more thought leadership.”

Panel result	Ingredient to touch first	Do not touch first
All four blank	Crawl / WAF	New cluster
Perplexity cites Reddit + a rival roundup; ChatGPT names the same rival	Corroboration	Copilot workshop
Copilot cites a Bing-indexed stale URL of yours	Recency	Training-bot debate
ChatGPT is accurate only when search is off	Residue vs live — force search in the panel	Celebrate the memory hit
Claude with web on invents a year	Packet + leftover URLs	Disallow ClaudeBot as the whole fix

That loop is the recipe under load. It is not a ranking-factor hunt.

What is not a ranking factor

Folklore fills the vacuum vendors leave. Kill it in the brief so nobody buys it.

Folklore	What is actually true	Source class
“Secret ChatGPT ranking factors”	Unpublished. Eligibility + extractability + corroboration + freshness are what operators can observe	Vendor docs do not list a score
FAQ schema forces recommendations	FAQPage JSON-LD does not mint citations. Extractable Q&A in HTML might get lifted	No vendor “FAQ = cite” rule
llms.txt is a citation switch	Optional hint file. Not a Search Console for ChatGPT	Not in OpenAI’s search-inclusion wording
Paying for ads buys chat citations	OpenAI documents OAI-AdsBot for ad landing-page safety, not foundation training, not a citation shortcut (OpenAI crawlers)	Ads ≠ answers
Google #1 transfers to ChatGPT	Semrush’s SaaS sample: 2.1% overlap with Google top 10	Small sample; still directional
Blocking all “AI bots” is privacy-complete	It often opts you out of search answers	Token split above
Date bumps = freshness	Engines learn empty lastmod	IndexNow FAQ: ping ≠ index
One engine win = all engines	Ahrefs: 86% of top sources unshared across three surfaces	Different pools

If a vendor or an agency sells you a numbered “AI ranking factor” list with no primary source, it is a pitch. Keep the recipe. Drop the numerology.

Failure mode: block-AI robots.txt plus a slogan homepage

This is the failure I see when a site looks “done” and still never gets named.

A 2023 security pass dumped every AI user-agent into Disallow. A later CDN managed list named “AI scrapers” and quietly included OAI-SearchBot and PerplexityBot. The homepage hero says “crafting exceptional experiences.” About lists two legal names. GBP, the footer, and schema disagree on the city. Pricing is a “book a call.” The blog calendar is ambitious. ChatGPT names three competitors with directory pages and a criteria table. Perplexity footnotes Reddit and a roundup. Copilot cites a Bing-indexed competitor. Claude, with web on, never fetches you.

Layer	What broke	Cost	Instead
Crawl	Search tokens Disallowed or challenged	Zero retrieval	Split training vs search; allow citation bots
Extract	No category sentence in HTML	Retrieved, unusable	60–80 word lead + matching schema
Corroboration	Only owned URLs	Unsafe to shortlist	Directories, reviews, one honest roundup
Recency	Old package still live	Wrong recommendation or skip	Ship the real offer; 301 the corpse
Measurement	One Slack screenshot	False win / false panic	Frozen panel, four columns

What it costs: you keep paying for content that cannot be fetched or cannot be defended. Competitors collect the shortlist. Sales hears “ChatGPT told me to call them” and you hear “AI doesn’t work for us.”

Do not start with a new cluster. Start with fetch logs and the entity packet.

How do you measure whether recommendations are working?

Citation rate is non-deterministic. Log a fixed 25–40 prompt panel. Re-run it. Treat one screenshot as anecdote.

Column	Values	Win
Engine	ChatGPT (search on), Perplexity, Copilot (web on), Claude (web on)	You actually used the live path
Prompt	Frozen text	Same string every week
Named?	Yes / no	Recommendation prompts need the name
Cited?	URL / none	Citation without a name is progress, not the buyer win
Accurate?	Match / stale / invented	Stale is a recency ticket
Sources	Domains in the panel	Corroboration map

Microsoft’s Bing Webmaster AI Performance report covers Copilot, Bing AI summaries, and “select partner integrations.” It does not name ChatGPT as a row (Bing AI Performance). Use it for Copilot / Bing AI. Do not paste it into a ChatGPT slide.

Week	What you record	Pass
0	Token table; entity packet; 30 prompts	Written, not vibes
1	Fetch 200s for citation bots on money URLs	Logs, not a robots.txt screenshot
2	Panel across four engines	Named / cited / absent / hallucinated
4	Same prompts	Direction, not a one-off

Prompt shapes that force a recommendation, not a definition:

Type	Shape	Recipe miss if…
Hire	“Who should I hire for {category} in {geo}?”	Rivals named; you absent
Best-of	“Best {category} for {ICP}”	Roundups cited; you unlisted
Alternative	“Alternatives to {incumbent}”	You are not in the set
Brand	“Is {legal name} a good fit for {job}?”	Invented year or offer
Compare	“{You} vs {rival}”	Rival’s table quoted; your URL never appears

Sample log row (copy the columns; do not invent a composite “AI score”):

Date	Engine	Named	Cited URL	Accurate	Top third-party sources	Ticket
2026-08-09	ChatGPT search on	No	—	n/a	rival roundup, Wikipedia	Corroboration
2026-08-09	Perplexity	No	—	n/a	Reddit, rival docs	Corroboration
2026-08-09	Copilot web on	Yes	/pricing (old)	Stale	Bing snippet of /pricing	Recency
2026-08-09	Claude web on	No	—	Invented year on brand probe	none	Packet + fetch

If ChatGPT and Perplexity both go dark the week you shipped an “AI” WAF, start with tokens, not with new posts. If fetch is healthy and you are still unnamed, you have a corroboration or extractability ticket — the competitors spoke’s ladder, not a fourth blog tool.

What to skip if you only have a week

A week cannot earn a roundup graph. It can make you eligible.

1. Diff robots.txt and the WAF managed list against the citation tokens above. Allow them on public money URLs.
2. Write the entity packet into About + homepage + one offer URL + Organization JSON-LD. Same strings.
3. Confirm Bing indexed the offer URL if Copilot or ChatGPT Search matters. IndexNow ping after the edit. Do not wait for a citation miracle.
4. Run 15 frozen prompts across the four engines. Screenshot sources. That is the baseline.
5. Pick one third-party fix you can actually reach (GBP category, a directory NAP, a review that names the job).

Skip:

Skip this week	Why
A 12-post calendar	Ingredient 2 and 3 are not “more URLs”
llms.txt as the project	Not a citation switch
A Copilot-only content program	Spot-check after Bing eligibility
Reddit spam	Corroboration is earned; spam is residue
Buying an “AEO score”	The panel is the scoreboard
Rewriting a passage that already gets quoted	Recency is accuracy

If the panel is still zero after fetch and the packet, stop and run the exclusion diagnostic. More publishing will not invent a trail.

When this is not worth doing yet

The recipe assumes a public, indexable business with a named category. Some teams are not there.

Situation	Do the recipe?	Do this instead
Offer is login-only / no public URL	Not yet	One public, quoteable page for the category job
Legal requires blocking all fetchers	Only if you accept invisibility	Written tradeoff; do not expect citations
You cannot name the category in one noun	Not yet	Positioning before AEO
No buyer uses these four products	Thin	Confirm in CRM; maybe Google-only is honest
Brand facts are in a lawsuit / rename	Pause recommendations	Finish the packet; then panel
You want a guaranteed citation date	Never	There is no SLA

If nobody on the sales team has ever heard a buyer mention ChatGPT, Perplexity, Copilot, or Claude, still keep crawl hygiene cheap — blocking search bots is hard to undo in a panic — but do not staff a four-engine program. Instrumented neglect beats a fake roadmap.

When it is worth it: high-consideration offers, inbound that already arrives as “I asked ChatGPT,” or a competitor set that is winning the shortlist while you own Google. Then the recipe is the work. The playbook is the system around it.

FAQ

What makes ChatGPT, Perplexity, Copilot, and Claude recommend a site?

They recommend a site they can crawl, extract as consistent entity facts, corroborate on other domains, and treat as current. Search bots and training bots are different tokens, so a training opt-out is not a citation strategy. Vendors do not publish a ranking score; eligibility plus a quoteable trail is what you can actually ship.

How do I measure whether makes ChatGPT, Perplexity, Copilot, and Claude recommend a site is working?

Run a frozen 25–40 prompt panel on ChatGPT with search on, Perplexity, Copilot with web on, and Claude with web on, and log named / cited / absent / hallucinated plus source domains. Repeat the same strings; one screenshot is anecdote. Bing’s AI Performance report can inform Copilot and Bing AI — it is not a ChatGPT console.

What usually fails first when teams try this?

Crawl access. A “block AI” robots.txt or a CDN bundle that challenges OAI-SearchBot, PerplexityBot, bingbot, or Claude’s search/user agents means the rest of the recipe never runs. Next failures are a slogan homepage with no category sentence, then a missing third-party trail. Fix fetch before you hire more writers.

How long does this take to show results?

Robots.txt adjustments are quoted around 24 hours by OpenAI and Perplexity; that is eligibility lag, not a recommendation SLA. Google recrawl can take days to months. IndexNow does not guarantee indexing. Corroboration and training residue take longer than a fetch fix — think weeks for a cleaner panel, not overnight.

What should I skip if I only have a week?

Skip the blog calendar, llms.txt theater, and a Copilot-only program. Allow citation crawlers, ship one entity packet onto About and a money URL, confirm Bing on that URL, and baseline 15 prompts. One directory or review fix you can actually reach beats twelve new posts nobody can quote.

When is this not worth doing yet?

When you have no public quoteable page, when legal blocks every fetcher, or when you cannot name a category noun. Also skip a four-engine program if buyers never use these products — keep cheap crawl hygiene, measure Google, and do not staff theater. There is no paid switch that forces a recommendation date.

CTA

If you want to be named in ChatGPT, Perplexity, Copilot, or Claude, ship the recipe — crawl, facts, corroboration, recency — before you buy another content calendar.

Lane: /visibility · Book a visibility audit.