Does original research help me get cited by AI

Yes — original research helps you get cited by AI when the finding is yours, unique, and extractable: a number or named result sitting next to methods, sample, and a collection window. No — it does not help when the page is an uncited blog opinion, a restated industry average, or a PDF nobody can quote. Models reuse facts they can defend. They do not award citations because you labeled a post “study.”

This is the decision spoke under the Answer Engine Optimization playbook. The publishing recipe lives on original research for AI citations. I have been SEO certified since 2021. The AEO version of that work is still one fact, one source, one URL. I will not invent a citation-lift percentage for “doing research.”

The short answer

• Yes when the page ships a unique, HTML-text fact with methods, sample, and date on the same screen
• No when the page is an uncited opinion, a recycled average, or a chart with no number a crawler can copy
• Reuse is credit assignment. Ambiguous claims get averaged. Unique measurements get attributed — if they are retrieved
• Ahrefs’ 75,000-brand analysis is a correlation (branded web mentions 0.664 Spearman vs 0.218 for backlinks), not a research ROI forecast
• Measure inclusion on a frozen prompt panel. Do not sell a lift % you cannot source

Does original research help — the yes/no gate

It helps when the research produces a citeable unit: a sentence a model can lift without inventing provenance. It fails when the research produces vibes. Google’s helpful-content questions still ask whether the page provides original information, reporting, research, or analysis. Their AI-optimization guide tells you not to recycle what others already said, or what a model could invent. That is a quality instruction, not a guarantee that your survey will appear in ChatGPT.

Google’s AI features page is the eligibility floor: the URL must be indexed and snippet-eligible. Important content has to exist as text. There is no extra “research markup” that forces a citation.

If this is true	Verdict	Why a model would bother
You measured something nobody else published, with n, population, and date	Yes, candidate	Credit is assignable
You restated “most marketers struggle” with no dataset	No	Credit is shared / none
You published a unique number inside a 40–80 word HTML block	Yes, if retrieved	Extractable span
The number lives only in a slide or a gated PDF	No	Nothing to quote
Five agencies already published the same rounded figure	No	You are a restatement
You invented respondents or dressed a customer list as a census	Toxic	Spreads, then poisons trust

The gate is not “did we do research.” It is “did we ship a fact the open web can reuse without guessing who owns it.”

• The headline claim is a fact, not a mood
• Only your URL can claim that exact measurement
• A stranger can copy the sentence without opening an image
• You would still publish it if AI products disappeared tomorrow

What kind of research actually gets reused

Answer engines do not “respect studies.” They reuse passages. The research types that survive compression are the ones that leave a clean, attributable residue: a unique number, a method clause that qualifies the number, and a sample clause that stops the model from promoting you into a national census.

OpenAI’s ChatGPT Search help is the operator check on the chat side: inline citations and a Sources panel appear when search ran. No Sources control means the answer came from memory, not a live page. Research you never indexed cannot win that round. Memory residue is a different clock.

Residue type	What gets reused	What gets stripped
Unique number	“41% of 87 B2B SaaS marketers (Mar 2026) said X”	“Most teams struggle with AI search”
Methods	“Email survey of current customers, opt-in, Mar 3–14 2026”	“According to our research” in the hero
Sample	n, who was invited, who was excluded	“Hundreds of marketers” with no n
Window	Collection dates next to the figure	“Recent data shows”
Negative finding	“We did not see X in this sample”	A buried caveat on page 14
Named comparison you measured	“Tool A vs B on the same 12 jobs”	A logo grid with no test

If a competitor could write the same sentence without stealing, you do not have research. You have a take.

Reuse prefers boring honesty over cinematic decks. A 40-row customer table with limitations will get quoted more often than a “State of the Industry” PDF with no n. The PDF looks expensive. The table is extractable.

Priority if you can only mint one asset this quarter:

1. A unique number you already have, HTML, methods on the same URL.
2. A named comparison you actually ran on a frozen job set — n jobs, same inputs, dated window.
3. A negative finding with the same disclosure (what you looked for and did not see).
4. A customer quote that names a measurable outcome, sitting next to the number — not instead of it.
5. Not: an annual twin of a vendor “state of” report with the same three round numbers.

schema.org Dataset and Google’s Dataset structured data can describe a real dataset landing page. They are not a citation cheat code. Google’s AI-optimization guide says structured data is not required for generative AI features. Markup that disagrees with the visible table is worse than no markup.

Should you compete with Pew, Gartner, or vendor annuals?

Usually no. Institutional research already owns the generic “how many adults use chatbots” sentence. You will not displace Pew by running 40 emails. You might get cited on a slice they will never run: your category, your jobs, your customers, your failure modes. That is the point of SMB-scale original work.

Question shape	Compete with institutions?	Better move
National incidence (“what % of US adults…”)	No	Cite them; add your slice
Category-specific (“what % of HVAC owners…”)	Maybe, if you have access	Your customer / job sample
Tool bake-off on public marketing copy	No	Same 12 jobs, same inputs, dated
“State of AI search 2026” clone	No	One question the annuals skip
Your time-to-X on a named workflow	Yes — they cannot copy the rows	Publish the window and n

Decision rule: if the sentence would still be true after you delete your brand name, it is not your research. If deleting your brand name makes the sentence impossible, you have a fingerprint.

• The prompt you care about is not already answered by a census-quality source
• Your access (customers, jobs, logs) is the scarce input
• You can live with a scoped claim (“in this sample”) in the public lead
• You are not trying to “beat Gartner” for a board slide

Beating a census is vanity. Owning a slice is AEO.

Extractable unique facts vs uncited opinions

Extractable means a retriever can lift a self-contained block and a synthesizer can attach a source without writing glue. Uncited opinion is the opposite: a confident paragraph with no owner, no n, and no date. Models can paraphrase opinion all day. They have no reason to name you when they do.

Google’s AI-optimization guide contrasts non-commodity content (unique experience, unique data) with commodity roundups anyone could generate. A first-hand measurement is non-commodity. “7 ways to get cited by AI” with no original figure is commodity, even if you spent a week on the adjectives.

Page shape	Extractable?	Typical citation behavior
Lead sentence = result + n + population + date	Yes	Cite the research URL if retrieved
Opinion essay, sources “the industry”	No	Restate, cite a roundup or nobody
Chart-only “insight,” number not in HTML	No	Skip, or guess from alt text and get it wrong
Quote from a named customer next to the number	Yes, as color	Quote + statistic travel together
“We believe brands must…” with no measurement	No	Treat as marketing copy
FAQ that repeats the number in 2–4 sentences	Yes	FAQPage-shaped spans are easy to lift

Run the copy-paste test before you brief a designer.

1. Highlight the headline finding.
2. Paste it into a blank note with no surrounding essay.
3. Ask: can a stranger tell what was measured, on whom, when, and by whom?
4. If any of those four is missing, the fact is not extractable.
5. If the sentence could appear on five other blogs this week, it is an opinion wearing a number-shaped hat.

Opinions can still rank. They rarely become the source of record for a generated answer. Credit assignment is the job. Uncited opinions have none.

Unique numbers: what “yours” actually means

“Yours” means you collected or computed it, you can describe how, and no other domain already owns that exact figure. A unique number is not “we rounded Gartner.” It is not “our AI estimated.” It is a measurement with a fingerprint: window, n, population, and a URL that is the canonical home.

Ahrefs’ brand-visibility work is useful here as a mention mechanic, not as a lift %. In their 75,000-brand analysis, branded web mentions correlated more strongly with AI Overview brand visibility than backlinks did (0.664 vs 0.218 Spearman). That is still a correlation. It does not say “publish a survey and gain X% citations.” It does explain why a number other people repeat — with your name on the sentence — is more useful than a homepage backlink with no sentence attached.

Number you might ship	Unique?	What a model should do
Your March 2026 customer survey, n=40, methods below	Yes	Cite you if retrieved
“72% of marketers” copied from a vendor slide	No	Cite the original vendor, or drop it
Internal win-rate from 12 closed deals, labeled as that sample	Yes, scoped	Cite with the limitation
Model-generated “industry average” with no dataset	No / toxic	Ignore, or launder a hallucination
Same unique figure reprinted on a trade site with your URL	Yes + corroborator	Cite either URL; both teach the fact

If you cannot name the rows, you do not own the number. If you own the number and hide the rows’ description, you are asking the model to invent the provenance. Hide less.

A unique number that never leaves your blog is still better than a fake one. It is weaker than a unique number a journalist can paste. Mentions are the distribution layer. The fact has to exist first.

Percentages are not always the right shape. Counts, ranges, and “did / did not” findings reuse just as well — sometimes better — because they are harder to round into someone else’s average.

Shape	When it is the right residue	When it backfires
Percentage	n is public in the same sentence	Naked “41%” with no n
Count	Small n, honest (“12 of 40”)	“Hundreds” as a mood
Range / windowed median	Messy real data	Fake precision (±1% on opt-in)
Binary / negative	You looked and did not find X	Burying the miss
Ranked list you measured	Same rubric, dated	A listicle with no scoring

Decision: pick the shape that survives a skeptical buyer, then put that shape in HTML. Do not pick the shape that looks like a McKinsey slide.

• The number cannot be googled as a vendor talking point
• The unit (% vs count) matches the sample
• Rounding is declared (41%, not 41.273%)
• The canonical URL is the research page, not a landing-page teaser

Methods: the provenance a model can defend

Methods are not academic theater. They are the clause that keeps the synthesizer from promoting a convenience sample into a law of nature. AAPOR’s disclosure standards exist for the same reason even if you are not a polling firm: sample size, how the sample was generated, and dates of data collection belong next to the result. Hide those and you are asking a model to invent the provenance.

Google’s helpful-content “How” questions are the same instinct: tell people how the work was produced. For a survey, that is recruit path, window, and what you cannot conclude. For a benchmark, that is the jobs, the tools, and the machine. For a content audit, that is the query set and the scoring rule.

Methods field	Why it gets reused	What “we researched this” hides
Collection window	Stops an aged number from traveling as current	“Recent”
Population	Stops “customers” from becoming “the market”	“Industry leaders”
How people got in	Opt-in vs panel vs census	“Statistically significant” with no model
What you did not measure	Negative space a model can keep	A footnote the extractor drops
Who computed it	Brand + person, not “our AI”	Anonymous dashboard export

Put methods on the same screen as the number. A methods appendix three scrolls down gets stripped. A model that only sees the hero will either drop you or cite you for a cleaner claim than you made.

• Window is a pair of dates, not “Q2”
• Population is a noun (customers, US dentists, n8n Cloud workspaces), not “stakeholders”
• Recruit method is one sentence
• Limitation is one sentence, not a PDF
• The same fields would still be true if a reporter called you

Methods are how original research stays original after compression. Without them, your unique number becomes everyone else’s unsourced average.

Sample: when a small n still counts

Small n counts when you label it. It fails when you inflate it. A 40-respondent customer survey is original research. Pretending those 40 people are “the American marketer” is not research. It is a press release.

SMB teams stall here because they think research means a probability sample. It does not. It means a described sample. AAPOR still wants size and recruit method next to the result on a convenience list. An opt-in customer list is fine. A “±3%” on that list, with no model, is theater.

Sample	Honest reuse	Dishonest reuse
n=40 current customers, email opt-in	“In this customer sample…”	“40% of companies nationwide…”
n=12 jobs on the same golden set	“On these 12 tasks…”	“Benchmarks show we win”
n=1,200 panel, vendor-recruited	Cite the panel house and weights	Hide the vendor
n=unknown “inbox of replies”	Do not publish a %	A viral round number
AI-generated rows mixed with humans	Do not publish	A fake census

The reuse test for sample is simple: would you say the sentence out loud to a skeptical buyer without wincing? If you would add “of our customers” in the room, put those three words on the page. Models copy the page, not the room.

1. Write the number.
2. Write the n in the same sentence.
3. Write who was invited.
4. Write who was excluded.
5. If step 3 or 4 is embarrassing, you do not have a publication yet. You have a huddle.

Do not publish yet if any of these are true:

Block	Why it fails reuse
You will not print n next to the %	The model will invent a census
Legal strips the population noun	Credit becomes “a survey”
Half the rows are duplicates or bots	The fingerprint is fake
You plan to announce, then “share methods later”	Later never ships

I will not tell you a minimum n that “unlocks citations.” There is no such threshold in Google’s public docs, and I will not invent one. Honesty scales down. Fabrication does not scale at all.

When a large study still gets ignored

A 2,000-row study can lose to a 40-row table. Size is not the citation. Extractability, uniqueness, and retrieval are. Large studies get ignored when they are gated, image-only, undated, already commoditized, or published on a URL that is noindex, nosnippet, or orphaned.

Google’s AI-features technical bar does not care that you hired a research firm. The page still has to be indexed, snippet-eligible, and textual. ChatGPT Search will not cite a login wall it cannot fetch. A press-only embargo PDF is a donor organ for whoever HTML-ifies the number first — often a journalist, not you.

Large-study failure	What you observe	First fix
Gated PDF / email wall	Trade sites get cited; you do not	Public HTML canonical
Number only in a chart	Overview cites a blog that retyped it	Put the figure in text
Methods behind “request the full report”	Model cites a roundup that rounded you	Methods on the same URL
Five vendors already run the same annual survey	Your “State of X 2026” is a twin	Measure a slice they do not
Canonical is /lp/ebook with nosnippet	Rankings maybe; citations no	Snippet-eligible research URL
Brand name missing from the sentence	Mentions go to the category, not you	“Spurlock Studios [window] survey…”

If the only public sentence is “Download the report,” you did not publish research. You published a lead form. Lead forms do not get cited.

The expensive failure is the 80-page PDF that exists to impress a board. The cheap win is one HTML finding other people can steal with the URL attached. Stealable-with-attribution is the point.

Failure mode: thought-leadership PDFs

The failure that burns the most money is the uncited “thought leadership” program: a designer, a survey vendor, a gated PDF, a LinkedIn carousel of adjectives, and zero extractable facts. Six weeks later the prompt panel still names competitors. Leadership concludes “original research does not work.” What failed was publication, not the idea of measuring something.

I have watched this adjacent to 500+ live automations and hundreds of production sites: teams confuse production cost with citeable residue. Cost is not a ranking factor. A $40k PDF with no HTML number loses to a $0 table a practitioner can paste.

Symptom	Cost	What to do instead
“We published a study” and no URL returns the number	Designer + vendor + zero citations	One public finding page
Chart thumbnails on social, number not on the site	Social reach, no retrieval	HTML first, creative second
Hero claim disagrees with the methods PDF	Model cites neither, or the rounder third party	One screen, one claim
Invented n to “look national”	Legal + trust when checked	Kill the draft
Refresh the date, not the data	Helpful-content trap	New window or take it down

Google’s helpful-content list is explicit about changing dates to look fresh when the content has not changed. An undated 2024 survey presented as 2026 truth is worse than no survey. Aged residue is how hallucinations get a costume.

Thought leadership without a fingerprint is an opinion with a budget. Do not fund it as AEO.

A second failure sits next to the PDF: the restated average with a new cover. Someone copies three public figures, writes 1,800 words of commentary, and files it as original research. Google’s helpful-content questions ask whether you provided original information or merely rewrote other sources. Commentary can be useful for humans. It is not a unique fact a model has to attribute to you.

You shipped	Original?	Likely model move
New measurement, your rows	Yes	Cite you if retrieved
Commentary on Pew / Gartner with full citation	Analysis, not your number	Cite the institution
Three public stats, no new n, “our take”	Opinion	Cite a roundup
Same stats, chart restyled, methods omitted	No	Skip or cite whoever HTML’d first

If the only original thing is the cover illustration, you did not do research. You did production design.

What a week can actually do

A week cannot run a defensible national study. A week can decide whether research is even the bottleneck, and it can ship one extractable fact you already own. If you do not already have rows, do not commission a panel this week. Fix the pages models can already quote.

The how-to spoke is the 14-day recipe. This page is the triage. Most teams who ask “does research help” this week actually have an extractability or corroboration problem, not a data problem.

This week	Do	Skip
You already have a unique internal number	Publish it with n, window, limitation	A new survey RFP
You have only opinions and restated averages	Do not fake a study	The PDF
Existing posts bury facts in essays	Pull one number into a lead + table	A new cluster of 12 blogs
Research URL is gated or nosnippet	Ungate the finding; restore snippets	Another ebook
Prompt panel already cites a competitor’s table	Match their extractable shape on a true unique fact	Adjectives

Checklist if you only have a week:

• Baseline 10–25 revenue prompts. Log mention vs citation. Method: measuring AI search visibility
• List facts you already own (win-rates, time-to-X, ticket mixes) that a competitor cannot copy without lying
• Ship one public HTML finding or ship nothing — do not ship a teaser
• Do not hire a survey vendor in five business days and call it science
• Do not invent n to look bigger than the week

A week of honesty beats a quarter of theater. If the honest version is “we have no unique number yet,” that is the finding. Spend the week on entity facts and extractable answers, not a fake census.

When original research is not worth doing yet

Skip original research when the site cannot be cited for any reason, when you have no unique question, or when nobody will log whether the number got used. Research is a citeable asset. It is not a crawl fix, not a NAP fix, and not a substitute for a prompt panel.

Blocker	Why research waits	What to do first
noindex, nosnippet, login wall on the would-be URL	Nothing to retrieve	Eligibility
Entity name / category / offer still in flux	You will cite last year’s product	Fact packet
You cannot name the question the number answers	Retrieval has no query to match	Pick one buyer prompt
Nobody will re-run the same prompts	You cannot tell if it helped	Measurement ritual
The “study” would restate a public average	You will be the fifth restatement	Find a slice they skipped
Legal will not let methods go public	The number will travel naked	Don’t publish, or publish the limitation

If Maps, crawl, or entity consistency is on fire, wait. Sequencing beats a new dataset on a broken host. The playbook’s truth layer still comes first.

• The research URL would return 200 to an anonymous crawler
• You can state the buyer question in one line
• You can describe the sample without a lawyer rewriting it into mush
• Someone owns a 30-day prompt log
• You are not using “research” to dodge a technical ticket

Not worth it yet is a valid output. Write it down. Do not fill the gap with an opinion labeled as a study.

What pages should you fix first instead of a study?

If the prompt panel already fails you, research is often the wrong first ticket. Fix the pages a model can already retrieve before you mint a new dataset. The target query “what pages should I fix first for AI search?” is usually an extractability and entity problem, not a missing 40-row survey.

Page class	Why it is first	Research waits until
Category / “who we are” entity page	Name, offer, NAP, same facts everywhere	The entity string is stable
One answer page for the money prompt	Retrieval needs a query-shaped URL	You know which prompt
Comparison / criteria page with a table	Models lift tables	You stop burying the answer in a manifesto
Existing post that already has a unique number buried	Cheap extractability win	The number is in the lead
Research URL (new)	Only after the above, and only if you own unique rows	Eligibility is clean

Procedure when leadership wants a study this month:

1. Dump the last 15 lost deals or inbound notes. Circle every “ChatGPT said / Google said / Perplexity said.”
2. Map those utterances to URLs you already have.
3. Quote-test the lead: can you paste 60 words that stand alone?
4. If the lead fails, rewrite that URL this week. Do not open a survey RFP.
5. If the lead passes and the panel still cites a competitor’s table of unique numbers, then research is on the table.

Pages to skip first: the 12th listicle, the gated ebook, the “insights hub” with no n. Those are opinions with a CMS.

• Entity page agrees with schema and the footer
• Money prompt has an HTML answer in the first screen
• At least one table or numbered procedure exists on that URL
• You have not used “we need original research” to skip that rewrite

What this costs in attention — not a lift percentage

I will not invent a citation-lift % or a payback multiple for “doing research.” The honest cost is attention and delay: designer hours, legal review, a URL you have to keep honest when the window ages, and a prompt log somebody has to run. A blog calendar of commodity posts costs attention too. It just hides the waste.

Spend	What you actually buy	When it is rational
One public finding page	A citeable unit	You already own unique rows
Survey vendor + gated PDF	A lead form with a chart	Almost never, for AEO
12 commodity blogs	Index bloat	Never, as a substitute for a number
Trade pitch of an existing number	Mentions (Ahrefs’ correlation is about mentions, not your %)	After the HTML fact exists
Visibility audit	A gate: research vs rewrite vs eligibility	The room disagrees

If you cannot staff the log, you cannot tell whether the asset worked. Then the research spend is a brand exercise. Call it that. Do not file it under AEO.

Google’s helpful-content questions still punish mass-produced pages that summarize other people. Twelve restated blogs are not “a research program.” They are the commodity pattern the AI-optimization guide told you to stop.

How you know it is working

You know it is working when your research URL (or a third-party page that repeats your number and names you) appears as a citation or a supporting link on a frozen prompt panel — not when the PDF download count goes up. Rank tracking is necessary and incomplete. Inclusion is the KPI.

Do not invent a share-of-voice percentage from one screenshot. Do not convert a lab paper’s historical visibility metric into your forecast. The how-to spoke covers the GEO paper and what it actually measured. This page will not recast that lab number as your lift.

Signal	Counts as working	Does not count
Research URL in ChatGPT Sources / inline cite on a frozen prompt	Yes	A memory answer with no Sources
Google AI Overview supporting link to the finding page	Yes	Impression with no idea which URL
Trade article repeats n + window + your name	Yes (corroboration)	A paraphrase with the number rounded off
GSC generative AI impressions on the research URL	Supporting, Google-only	ChatGPT / Perplexity proof
PDF downloads, likes, “thought leadership” mentions	Distribution maybe	Citation

Log mention and citation as two columns. A model can use your number and drop your URL. That is a mention. It is not the win you paid for, but it is not zero. The measurement spoke is the scoreboard. Re-run the same prompts. One lucky day is an anecdote.

1. Freeze the prompts before you publish.
2. Publish the finding.
3. Re-run on a fixed cadence (day 0 / 7 / 14 / 30 is enough to start).
4. Record cited URL, not “we were in the answer somewhere.”
5. If the number shows up without you, you have a corroboration / attribution job, not a new survey.

This page is the gate. The April spoke is the recipe. Mixing them is how a team commissions a survey before they can publish a sentence.

Job	This page	Original research for AI citations
Question	Does it help, and what residue gets reused?	How do I publish a number a model can extract?
Output	Yes / no / not yet + asset type	Methods block, Dataset notes, seeding, refresh cadence
GEO paper	Point, do not recast as your %	Full table and hedges
Measurement	Inclusion of the research URL	Panel plus PR interaction
CTA	Audit if the gate is ambiguous	Same offer after you decide to ship a number

If the verdict is yes, open the how-to and ship one finding. If no, stop buying PDFs. If not yet, fix eligibility and extractability on pages you already have. The playbook remains the system: entity facts, machine-readable truth, passages that survive compression, off-site agreement, measurement. Research is one way to mint a passage. It is not the whole stack.

Decision checklist for the next meeting

Print this. Fill it in 20 minutes. If you cannot, you are not ready to fund a study.

• Question: Which buyer prompt should this number answer? Write it verbatim.
• Uniqueness: Could a competitor publish the same sentence this week without stealing rows?
• Residue: Will the public page contain number + n + population + window in HTML?
• Sample honesty: Can we say who was measured without inflating to “the market”?
• Methods: One screen, not a request form.
• Eligibility: Indexed, snippet-eligible, not a gate.
• Owner: Who updates the window when the data ages?
• Scoreboard: Which frozen prompts, logged where? See measuring AI search visibility.
• Kill rule: If the panel is unchanged at day 30, we inspect extractability and corroboration — we do not invent a lift % to defend the PDF.

Verdict language you can actually use:

Fill-in	Meaning
Yes — ship one finding	We own unique rows; publish HTML this cycle
Yes — after eligibility	The fact exists; the URL cannot be quoted yet
No — opinion, not research	We have takes. Do not dress them as a study
Not yet	Crawl, entity, or measurement ritual is the real ticket

A unique, extractable, method-labeled fact is one of the few assets an answer engine has a reason to attribute. An uncited blog opinion is content. Do not confuse the two. Do not buy a percentage to paper over the difference.

If the room still wants a study after this checklist fails, the product is theater. Fund it as brand. Do not file it under citations.

FAQ

Does original research help me get cited by AI?

Yes if it produces a unique, extractable fact with methods, sample, and a date a model can lift in one block. No if it is an uncited blog opinion, a restated average, or a gated PDF. Citation is retrieval plus credit assignment, not a prize for having a research vendor.

How do I measure whether original research is helping me get cited by AI?

Re-run a frozen prompt panel and log whether your research URL — or a third-party page that repeats your number and names you — appears as a citation. Treat Search Console generative AI impressions as Google-only supporting evidence. Do not invent a citation-lift percentage from one screenshot or from a lab paper on a different metric.

What usually fails first when teams try this?

Publication, not curiosity. The number never becomes HTML, methods get gated, or the claim is a recycled average with no owner. Eligibility failures (nosnippet, login walls) look like “research does not work” when the crawler never saw a sentence.

How long does this take to show results?

There is no research-specific SLA. Live retrieval can surface a newly indexed, quotable page in hours to days; crawl and snippet eligibility often take weeks; memory residue lags longer. Day 30 is an eligibility check. Day 90 is when citation-rate movement is defensible. I will not invent a faster clock.

What should I skip if I only have a week?

Skip commissioning a new survey, a gated ebook, and any invented n. Publish one fact you already own with methods, or spend the week on eligibility and extractability. A fake census in five days is worse than no study.

When is this not worth doing yet?

When the would-be URL cannot be retrieved, entity facts are still moving, you have no unique question, or nobody will log a prompt panel. Fix crawl, snippet eligibility, and the scoreboard first. Research cannot rescue a page the engine is not allowed to quote.

CTA

If you cannot tell whether you need a finding page or just an extractable rewrite, that is the audit.

Lane: /visibility · Book a visibility audit.