How to Audit a ChatGPT Shortlist in a Week

You audit a ChatGPT shortlist in a week with a locked prompt set, a daily log of who gets named, and one page left to fix on Day 7. You do not buy a "visibility score." You run the same buyer prompts for seven days, write down who ChatGPT names, note which URLs it cites when it cites anything, and leave Monday with one page worth fixing before you touch ads.

I am William Spurlock, AI Visibility Strategist and Fractional AI CTO at Spurlock Studios. I have spent 20,000+ hours architecting agentic systems, built 600+ automations with 500+ of them live, and that work has saved clients 35,000+ hours. None of those receipts matter to ChatGPT until they sit on a page a stranger can check. This spoke is the owner-week version of the mechanism I cover in how ChatGPT and Perplexity actually decide which businesses to recommend. The playbook for earning the slot lives in how to get ChatGPT and Perplexity to recommend your business. Here I am only answering how you measure the shortlist without turning the week into a research project.

As of September 2026, ChatGPT answers can still come from training memory, live browsing, or a mix, depending on the account and the query. OpenAI's own product pages describe browsing and citations as features that appear when search is available, not as a public ranking API for brands. Perplexity cites live sources on almost every answer. Treat the week as a field check, not a lab score.

What kind of content makes AI systems like ChatGPT trust your website? #

ChatGPT trusts pages it can defend: a named author, a dated claim, a method a peer could argue with, and at least one outside page that repeats the same fact. Soft brand copy fails that test. A slogan cannot be a citation. A FAQ sentence with a city, a service, and a year can.

I judge trust the way an answer engine has to judge it when a buyer asks "who should I hire." The model needs a sentence it can lift without the paragraph before it. If your homepage only says you are a leading provider, there is nothing to lift. If your service page says you handle emergency HVAC in Traverse City, name the after-hours window, and point to a review or local article with the same city and trade, the model has something to stand behind.

What "trustable" looks like on the page:

Content type	Why ChatGPT can use it	What usually fails
Direct-answer service page	One claim per section, plain text, city and scope in the first screen	Hero metaphors and "our team of experts" with no names
Method or process post	Steps, constraints, and when you would refuse the job	Ten tips that could sit on any competitor site
Proof page or case narrative	Dated detail, named operator, checkable outcome definition	Invented ROI, anonymous "client," no year
Author or About page	Same legal name and role as the byline and LinkedIn	Stock photo "team" and a nickname that never matches filings
Third-party corroboration	News, directory, or review that repeats your name next to a fact	Logo walls that never attach to a sentence

I keep a short trust checklist when I open a site for the first time:

1. Find a human name within two clicks of the homepage
2. Confirm the money page answers the buyer's question in the first two paragraphs
3. Source every hard number with a URL and date, or hedge it as an estimate
4. Check whether a competitor's page would look different if I swapped the logo
5. Look for any page I do not own that repeats the same brand + fact pair

That last item is the one owners skip. Your own site can be clean and still lose the shortlist if every other mention of you is a social caption with no durable URL. ChatGPT and Perplexity both prefer evidence that survives a stranger checking the citation. Google's creating-helpful-content guidance, last updated December 10, 2025 on Search Central, still frames trust as something systems notice through a mix of signals, not a single score you buy. The letters E-E-A-T are Google's. The behavior is any engine that has to attach a sentence to a URL.

Current models I use when I pressure-test copy for quotability are Claude Opus 4.8 and Claude Sonnet 5 on Anthropic's side, GPT-5.5 on OpenAI's side, and Gemini 3.1 Pro when I want a second retrieval style. The model name is not the strategy. The quotable page is.

Content shapes I refuse to ship when the goal is AI trust:

• Homepage carousels where the real offer only exists inside an image
• "Thought leadership" that never names a constraint or a refusal
• Comparison pages that trash competitors without primary sources
• Blog posts that open with industry weather and never answer a buyer question
• Schema spam that claims LocalBusiness fields the page text does not support

If you only fix one thing before the audit week, fix the page that should win the money query. Make it boring and checkable. Engines quote boring when boring is true.

A practical rewrite order when the money page is soft:

1. Replace the first two paragraphs with the buyer answer in plain sentences
2. Add the service area and the hours or response window as text, not icons only
3. Put the operator name and role where a human expects a byline or an About link
4. Attach one source URL and date next to any statistic you keep
5. Delete the generic tip stack that could live on any competitor domain

I do this before the audit week on purpose. If you audit first and rebuild later, you still learn who owns the shortlist, but you waste a week measuring a page you already know cannot be cited. If the site is already quotable, skip straight to the prompt set.

How does an owner audit a ChatGPT shortlist in a week? #

Lock ten buyer prompts, run them once a day in the same ChatGPT account settings, log every brand named and every URL cited, then spend the last day ranking gaps by revenue impact. That is the whole audit. No dashboard required.

I treat the shortlist as a weekly snapshot, not a lifetime ranking. Answers drift when browsing is on, when the prompt changes by one adjective, and when a competitor ships a better page. The point of seven days is to see the pattern through the noise.

Day 0 setup (Sunday night, 45 minutes) #

Write the prompt set before you open ChatGPT. If you invent prompts mid-week, you are measuring your curiosity, not the shortlist.

Slot	Prompt shape	Example pattern
1	Category + city	"Who are the best [trade] companies in [city]?"
2	Category + urgency	"Who should I call for emergency [service] near [city] tonight?"
3	Category + constraint	"Who handles [service] for [industry] without requiring a five-figure retainer?"
4	Comparison	"Compare the top [service] options in [city] for a [buyer type]."
5	"Alternatives to"	"What are alternatives to [market leader] for [job] in [region]?"
6	"Who do you recommend"	"Who do you recommend for [job] if I care about [proof type]?"
7	Price-adjacent	"Who is worth hiring for [job] if I need a clear scope before a deposit?"
8	Mistake-avoidance	"How do I avoid hiring a bad [trade] in [city]?"
9	Vendor shortlist	"Give me a shortlist of [service] providers for [use case]."
10	Your brand probe	"Tell me whether [Your Brand] fits [job] in [city], and name other options to consider."

Rules for the set:

• Use the words your buyers type, not your internal product names
• Keep city, service, and buyer type consistent across the week
• Do not include your brand in prompts 1 to 9. Save prompt 10 for the brand probe
• Save the exact strings in a sheet so Monday-you and Friday-you run the same test

Account settings matter. Note whether browsing or web search is on. Note the model shown in the UI (for me that is usually GPT-5.5 on paid ChatGPT as of September 2026). If you switch models mid-week, label the rows. A shortlist from a browsing session is not the same evidence as a shortlist from memory-only.

Days 1 to 5: run and log (20 to 30 minutes per day) #

Same time of day if you can. Same ten prompts. Fresh chat per day so prior answers do not steer the next ones.

Log columns I actually use:

Column	What to write
Date	YYYY-MM-DD
Prompt ID	1 to 10 from your locked set
Model / browse	Model name and whether search ran
Brands named	Exact names, in order if the answer ranks them
You appeared	Yes / no / partial (mentioned without a clear recommend)
Cited URLs	Any links shown, split by your domain vs competitor vs directory
Tone	Recommend, warn, "I cannot verify," or generic category advice
Screenshot path	One file per prompt so arguments later are not "I think it said"

What I watch for while logging:

1. The same three brands on prompts 1, 4, and 9. That is a concentrated shortlist
2. Directories and review sites cited more often than operator sites. Your page is not winning the citation even if your category is discussed
3. Your brand on prompt 10 only. Training memory may know you exist while recommendation prompts still ignore you
4. Soft hedges ("I cannot verify current availability") when competitors get hard names. Your entity file is thin
5. Wrong city, wrong service line, or a dead URL if any citation appears. That is a trust burn, not a win

If ChatGPT refuses to name businesses, do not "jailbreak" it. Note the refusal. Change one constraint in a separate scratch prompt after the official run if you need a second angle, and mark that row as off-protocol. The audit is worthless if you keep nudging until you hear your name.

Pass / fail signals I mark in the margin while the answers load:

• Pass signal: the same competitor set repeats across urgency, comparison, and "who do you recommend" prompts
• Pass signal: a cited URL matches a live page with the same city and service in the first screen
• Fail signal: your brand appears only after you mentioned it in an earlier chat the same day
• Fail signal: the answer invents a phone number, address, or award you cannot find on any URL
• Fail signal: two answers in the same day contradict each other on who is "best" with no citation change

Invented contact details are a hard fail for trust, even if your name shows up nearby. Screenshot that row. You will need it when someone on the team says "but it mentioned us."

Day 6: competitor and citation map (60 to 90 minutes) #

Pull the five-day log into one view.

Finding	What it usually means	Owner move next week
Competitor A on 7+ of 10 prompts	Strong off-site corroboration or a cleaner money page	Diff their service page and their top two third-party mentions
Review site cited, nobody's .com	Engines trust the directory more than operator pages	Earn one independent mention, then tighten your own quotable sentence
You appear with wrong facts	Entity collision or stale page	Fix the page first, then ask for a re-check next week
You never appear, even on prompt 10	Thin entity + thin corroboration	Do not buy ads yet. Build the proof stack below
You appear only when browsing is off	Memory signal without live page strength	Ship the direct-answer page and get one outside repeat

I also open Perplexity once with the same ten prompts on Day 6. I am not mixing engines in the daily ChatGPT log. I am checking whether the shortlist is ChatGPT-specific or category-wide. Perplexity's citations make the gap obvious: if it cites three competitor URLs and zero of yours, you already know the rebuild target.

Day 7: decide the one page (45 minutes) #

Owners lose the week by leaving with twelve priorities. Leave with one.

Decision order I use:

1. Highest-revenue query where you are missing from the shortlist
2. Page that should answer that query today (service page beats blog fluff)
3. Single sentence that should become the citation
4. One outside URL that could repeat that sentence within 30 days
5. Re-run date on the calendar (same ten prompts, one week later)

The deliverable from the week is a one-page memo, not a slide deck:

• Prompt set (exact strings)
• Appearance rate: prompts where you were named / 10, averaged across days
• Top three brands that own the shortlist
• Citation gap: how often a URL is shown, and whose domain wins
• The one page to fix, with the draft citation sentence written out
• What you will not do this month (usually: ads, a full site redesign, or "more content" with no owner)

That is how an owner audits a ChatGPT shortlist in a week. The mechanism essay is the pillar. This is the field notebook.

A short example of what "done" looks like on Day 7 for a fictional local trade (no real client named):

• Appearance rate: 1 / 10 averaged across five days (brand probe only)
• Shortlist owners: Competitor A, Competitor B, a national directory brand
• Citations: review site three times, Competitor A .com twice, your domain zero
• One page to fix: /emergency-[service]/ with a citation sentence that names city, after-hours window, and licensed operator
• Outside target: one local association or news URL that can repeat that sentence
• Not doing this month: ChatGPT ads, a homepage redesign, ten new blog posts

Your numbers will differ. The shape of the memo should not.

Should a business buy ChatGPT ads or earn the citation first? #

Earn the citation first unless you already win the organic shortlist and only need paid reach for a temporary campaign. Ads can buy attention. They do not repair a thin entity file. If ChatGPT will not recommend you from evidence, a sponsored slot teaches buyers a name the rest of the answer still undercuts.

I am blunt here because owners ask this after one scary competitor screenshot. OpenAI has shipped advertising experiments and product surfaces that change over time. Treat any "ChatGPT ads" pitch as a dated product claim and check OpenAI's current ads or business pages the week you buy. As of my September 2026 reads, paid placement is not a substitute for the corroboration pattern in the parent pillar. If a vendor promises that ad spend equals recommendation share, ask for the primary source dated in 2026. If they cannot show one, walk.

Decision table I use with founders:

Situation after your week audit	Ads first	What I do instead
You never appear on prompts 1 to 9	No	Fix the money page and earn one independent mention
You appear with wrong city or stale services	No	Correct the entity pages before any paid surface learns the error
You appear organic on 6+ prompts and need a launch spike	Maybe	Paid as a temporary layer on top of a clean citation base
Directories win every citation and your .com is invisible	No	Become quotable on your domain. Ads do not invent that URL
Sales asks for "AI awareness" with no offer page	No	Ship the offer page. Awareness without a citation sentence wastes spend

Why citation-first wins for most local and B2B service brands:

1. Recommendation answers mix memory, retrieval, and caution. A paid label does not make a weak About page trustworthy.
2. Buyers who ask ChatGPT for a shortlist often click the cited source or Google the named brands. Your owned page still has to close.
3. Wrong paid messaging locks a bad claim into more surfaces. Fix the claim once on the site.
4. Competitors who already own reviews and local news keep winning unpaid shortlists while you rent a banner.
5. Your week audit already told you who owns the slot. Copy their evidence pattern before you copy their media budget.

When I would consider paid after the citation base exists:

• You have a dated, named proof stack and a clean money page
• The week audit shows you on the shortlist for the core query
• The campaign has a hard stop date and a measured CTA on a page you control
• You are testing a new city or offer while the organic entity file catches up
• Legal and claims review already signed the ad copy against the same facts on the site

I do not invent spend guidance. I will not tell you a monthly number. I will tell you the order: quotable page, corroboration, then optional paid. If your audit week showed zero organic mentions, buying attention is how you fund someone else's shortlist story about the category leaders.

Questions I ask before anyone opens an ads dashboard:

1. Confirm we appeared on the unpaid shortlist for the money query at least once this week
2. Check that the landing URL is the same page we want cited, not a campaign microsite with thinner proof
3. Match the ad claim to a sentence already live on that URL with a date
4. Name who owns the organic shortlist today, and what evidence they have that we lack
5. Set the kill date if paid does not move the next audit week's appearance rate

If you cannot answer those five, you are not ready to buy. You are ready to finish the citation base.

What proof does a brand need before ChatGPT concentrates recommendations on it? #

ChatGPT concentrates recommendations when the same brand-fact pair shows up on your site and on pages you do not own, with enough consistency that the model can repeat the name without looking reckless. Concentration is not "we posted daily." It is corroboration density around one clear offer.

I split proof into four layers. Skip a layer and the shortlist stays soft or shared with whoever completed the stack.

Proof layer	Minimum I want before expecting concentration	Failure mode
Identity	Legal business name, operator name, same spelling on site, maps, and LinkedIn	DBA chaos, "our team," mismatched cities
Offer clarity	One primary service sentence a model can quote without the hero copy	Five services fighting for the same H1
On-site evidence	Dated method, scope, service area, and contact path in crawlable text	Facts trapped in PDFs or image-only graphics
Off-site repeats	At least one independent URL that repeats brand + fact in the same words	Only self-mentions and social posts without durable URLs

Concrete proof objects that move recommendation odds in my audits:

1. A service page that answers who / where / what / when in plain sentences
2. An About or author page that matches the byline and the LLC name
3. Reviews or case notes that name the service and the city without fluff
4. A local news, association, or directory page that is not a paid dump of identical blurbs
5. Consistent NAP and category labels across the profiles buyers already trust in your trade
6. A refusal note: when you are not the right fit. Models and buyers both trust specificity
7. Sources next to any statistic, with the URL and date in the sentence

What does not count as concentration proof:

• A logo bar of "As seen in" with no linked article
• Anonymous testimonials with round ROI percentages and no year
• Press releases only hosted on your domain
• Schema that claims awards the body text never mentions
• A podcast appearance whose show notes never name your company correctly

How I know the stack is "enough" to expect tighter shortlists:

• Your Day-7 re-run shows you on more than one prompt family, not only the brand probe
• Perplexity cites your URL at least once for the money query, or cites an independent page that names you correctly
• Competitors still appear (that is normal). You are no longer invisible
• Wrong-fact mentions drop after you corrected the entity pages
• Sales starts hearing "I saw you recommended in ChatGPT" without you prompting the prospect

I still will not promise a date when ChatGPT "concentrates" on you. Training updates, browsing mixes, and category competition all move. The proof stack is the controllable part. The week audit is how you know whether the stack is working.

A 30-day proof plan that fits after the audit week, without inventing spend:

Week	Owner focus	Exit check
1	Ship the one money-page rewrite from Day 7	Citation sentence live in HTML text
2	Align About, maps, and LinkedIn to the same legal name and city	No spelling splits across the three
3	Request or earn one independent repeat of brand + fact	URL saved in the audit sheet
4	Re-run the ten prompts for seven days again	Appearance rate and citation hosts compared to baseline

You can compress that plan if you already have the outside mention. You cannot skip the money page and jump to week 4. The re-run without the rewrite only tells you the shortlist is still someone else's.

If you want the mechanism behind why two engines shortlist differently, read the parent pillar on how ChatGPT and Perplexity actually decide which businesses to recommend. If you want the earn-the-slot tactics after this audit, use how to get ChatGPT and Perplexity to recommend your business.

FAQ #

How does Claude decide what businesses or websites to recommend? #

Claude does not run a public business-directory ranker. It answers from context, tools, and what it can justify. In Claude.ai or API setups with web tools enabled, Anthropic's models (I currently pressure-test with Claude Opus 4.8 and Claude Sonnet 5) still need quotable, consistent sources the same way ChatGPT does. Without browsing, recommendations lean on training knowledge and whatever you pasted into the chat. Treat Claude as a second shortlist check in your week, not as a separate "Claude SEO" product. Same proof stack: named pages, dated claims, outside corroboration.

How do I get my blog to show up as a source in Perplexity answers? #

Write the post as a direct answer with crawlable text, then earn at least one independent page that agrees with your core claim. Perplexity retrieves live sources and cites them inline. A blog that never states a checkable sentence loses to a thinner page that does. Put the answer in the first screen, keep the claim consistent with your service page, and avoid trapping the facts in images. After publishing, re-run your money query in Perplexity and read the citations. If directories sites win, study their sentence shape, then improve yours. Indexing and clear HTML still matter. There is no special Perplexity meta tag that replaces a quotable page.

How do I track when ChatGPT or Perplexity mentions my brand? #

Build a weekly prompt log first. Add tools only after the manual baseline exists. I start with the same ten prompts, screenshots, and a simple sheet: date, engine, prompt, brands named, URLs cited. Paid monitoring products exist and change often. Verify their methodology and date before you trust a "share of voice" number. Alerts that only watch your exact brand string miss the painful case: you are absent while three competitors own the category shortlist. Re-run the audit week monthly for the money queries. That tracking loop beats a vanity dashboard you never open.

Get an AI-visibility-ready site #

If your week audit showed a clean shortlist owned by someone else, the fix is usually not more posts. It is one quotable money page, a matching identity record, and the first outside mention that repeats the same fact. That is the site I build when the brief is AI visibility: direct answers, named proof, and pages an engine can cite without guessing.

I am William Spurlock at Spurlock Studios. If you want that rebuild scoped against your actual ChatGPT and Perplexity shortlist, start from the contact section on this site or email william@spurlockstudios.com with your city, your primary service, and the ten prompts you already ran.