ChatGPT included a source link in just 6.8% of its answers in May 2026, up from 1.6% a year earlier. For professional services questions, it was under 4%. So most of the time an AI names your category, nobody sees a link at all. You are either in the answer or you are not, and there is no report telling you which.
Every agency selling you AI visibility is selling tactics. Almost nobody is selling instrumentation. This guide gives you the instrumentation first: a ten-minute audit you can run right now, four metrics you can calculate in a spreadsheet without buying anything, and then the fixes in the order that actually pays.
Key takeaways
- Check logged out. Use an incognito or logged-out session on every platform. A signed-in ChatGPT personalises answers from your history and will show you yourself.
- Get a baseline in 10 minutes. Run five prompts on ChatGPT, Perplexity, Google AI Mode, Gemini and Copilot: three test whether you appear in your category, two test whether AI describes you correctly.
- Take your prompts from real buyers. Build 20 to 30 prompts from sales calls and enquiries, run each three times per platform, and repeat monthly. One run per prompt tells you almost nothing.
- Track four numbers, no tool needed. Citation rate, share of voice against a fixed competitor set, accuracy, and average position all come from one spreadsheet and about two hours a month.
- Wrong is worse than absent. If more than 20% of your mentions are wrong or partial, treat it as urgent. Find the source page behind the error in Perplexity's citations and fix that first.
- Fix in order, starting this week. Open crawler access and make your one-line description identical everywhere, then claim entity records and ship Organization schema, then build off-site mentions.
- When a competitor is named instead of you, check why. Most of the time they wrote the page that answers the buyer's question. Write a more specific one.
The 10-minute audit: is AI already recommending you?
You do not need a tool for the first pass. You need a private browser window and a place to write things down.
Use a logged-out or incognito session on every platform. ChatGPT personalises answers from memory and chat history, so an account that has discussed your own company forty times will show you yourself and you will conclude everything is fine. It is not a real reading.
The five prompts to run right now
Run all five on ChatGPT, Perplexity, Google AI Mode, Gemini and Copilot. Replace the bracketed text with your own category and city.
- Who are the best [your category] companies in [your city or country]?
- I need help with [the specific problem you solve]. Which companies should I talk to?
- Compare the top [your category] providers. What are the strengths and weaknesses of each?
- What is [your brand name]?
- Is [your brand name] any good? What do people say about them?
Prompts 1 to 3 tell you whether you exist in the market’s answer. Prompts 4 and 5 tell you whether the AI knows who you are and whether it likes you. Those are separate failures with separate fixes.
What to record
For each run, note five things: whether your brand appeared, which competitors appeared, in what order they were named, whether a source link was shown, and whether what the AI said about you was accurate.
Ten minutes, five prompts, five platforms. That is your baseline.
How to read Perplexity’s source list
Perplexity is the easiest platform to learn from because it shows its work: every claim carries a numbered source card, and the full list sits under the answer. Copilot shows an aggregated source panel and Google AI Mode shows links by default, while Gemini is inconsistent, so start where the evidence is clearest.
Read the list in three passes:
- Whose sources are these? If eight of ten cite one competitor’s blog, that competitor has become the category’s reference material. That is a content gap, not a ranking problem.
- What type of source? Forum threads and video carry real weight. Reddit accounted for 46.7% of Perplexity’s top-10 source share in one analysis of 680 million citations. If your category’s answers are built on sources like these, publishing another blog post will not move it.
- Is your own page in there, and which one? Being cited for your homepage and being cited for your pricing page mean very different things about how ready a buyer is.
What “cited but wrong” looks like
This is the failure mode people miss, because it looks like a win. Your brand appears, so you stop reading.
Look for four things in prompts 4 and 5:
- Stale positioning. It describes the service you led with three years ago.
- Wrong scope. It says you serve enterprises when you serve SMBs, or names a country you have never worked in.
- Invented specifics. Pricing you never published, team sizes you never claimed, awards you never won.
- Damning with faint praise. “A smaller agency, best suited to businesses with limited budgets” is a sentence that costs you deals, and it is technically a citation.
Score each mention as accurate, partially accurate, or wrong. A brand cited wrongly in 40% of answers has a bigger problem than a brand cited in 20% of answers correctly.
Turning the audit into an ongoing check
Five prompts get you a baseline. They will not get you a programme. The difference between a hobby check and a real measurement system is whether your prompts sound like your buyers or like a marketer imagining your buyers.
So do not invent them. Take them from material you already have: sales call recordings, the first message in your inbound enquiries, the objections your team answers weekly, and the exact phrasing buyers use to describe their problem.
The four prompt types
Build 20 to 30 prompts across four types.
Category prompts (about 8). The buyer knows the category and wants a shortlist. “Best SEO agencies for B2B SaaS.” These decide whether you make the consideration set.
Comparison prompts (about 6). The buyer has a shortlist and is narrowing. “X vs Y.” “Alternatives to [the market leader].” Name your real competitors here. If you are not in the answer to “alternatives to [leader],” you are invisible at the exact moment someone is shopping.
Problem prompts (about 8). The buyer does not know your category exists and describes a symptom instead. “Our organic traffic dropped after the last Google update, who can help?” These are the highest-intent prompts and the ones most brands never test, because they are the hardest to guess and the easiest to pull from real sales calls.
Brand prompts (about 4). “What is [brand]?” “Is [brand] legitimate?” “What does [brand] charge?” These measure accuracy, not presence.
How many runs, and how often
Given how much AI answers vary, one run per prompt tells you almost nothing. Run each prompt three times per platform in separate fresh sessions and record all three. That is 25 prompts by 3 runs by 5 platforms, or 375 data points, and at roughly 20 seconds each, about two hours of work.
Do it monthly. Not weekly, because the noise will swamp the signal and you will chase ghosts. Not quarterly, because these systems shift fast: in one 13-week study, the sources ChatGPT leaned on changed sharply inside a single month.
The spreadsheet
One row per prompt run, ten columns.
| Column | What goes in it |
|---|---|
| Date | Run date |
| Platform | ChatGPT, Perplexity, AI Mode, Gemini, Copilot |
| Prompt ID | P01 to P25, so you can compare across months |
| Prompt type | Category, Comparison, Problem, Brand |
| Brand mentioned? | 1 or 0 |
| Position in answer | 1 = first named, 2 = second, and so on. Blank if absent |
| Competitors named | Comma separated, in order |
| Sentiment | Positive, Neutral, Negative |
| Accuracy | Accurate, Partial, Wrong |
| Source cited | The URL, if one was shown |
Everything in the next section is a pivot table on this sheet. That is the whole system: two hours a month and nothing else.
What to measure: four metrics that mean something
“AI visibility” is not a metric. These four are, and all of them come off the sheet above with no tool.
1. Citation rate
Citation rate = (runs where your brand appeared ÷ total runs) × 100
Calculate it three ways: overall, by platform, and by prompt type. The split matters more than the headline number. A brand at 30% overall might be at 60% on brand prompts and 4% on problem prompts, which means AI knows you exist but never suggests you to someone describing your problem. That needs a completely different fix from a brand sitting at a flat 15% everywhere.
Ignore any tool that reports a “ranking position in AI.” Visibility percentage across many runs is the stable measure. Position on a single run is noise.
2. Share of voice against named competitors
Share of voice = (your mentions ÷ total mentions of you plus your named competitor set) × 100
Pick four to six real competitors at the start and keep the set fixed, because changing the set changes the number without anything changing in reality.
This is the metric that survives a board meeting, because it is relative. If your citation rate falls from 22% to 18% but your share of voice holds at 30%, the category got noisier and you held position. Citation rate alone would have read as a failure.
3. Sentiment and accuracy
Two separate scores on the same mention, and most teams collapse them into one by mistake. Sentiment asks whether the framing is positive, neutral or negative. Accuracy asks whether the facts are right.
Track accuracy as its own rate: wrong or partial mentions ÷ total mentions. Anything above 20% is urgent, and usually fixable in weeks, because inaccuracy almost always traces to one bad source or a missing source of truth.
4. Position within the answer
Where does your name appear: first, mid-list, or in a trailing “you might also consider” sentence? Record it as an integer, but never treat a single run as meaningful. Track average position across runs where you appeared and compare month over month. Movement from 4.2 to 2.8 across 75 runs is real. Movement on one prompt is not.
What a realistic baseline looks like
Most businesses running this for the first time land in one of these bands.
| Overall citation rate | What it means | What to do |
|---|---|---|
| 0 to 5% | You are effectively absent. Likely an access or consistency problem, not a content problem | Start at Week 1 below and do not skip it |
| 5 to 15% | Normal for a small or mid-sized brand with decent SEO. You appear when the question is narrow | Entity and structure work will move this fastest |
| 15 to 30% | Strong. You are in the consideration set for your core category | Focus on problem prompts and share of voice |
| 30%+ | Category leader territory | Protect it. Monitor accuracy and competitor displacement |
Two things to calibrate against. In the headphones category, SparkToro found Bose, Sony, Sennheiser and Apple appeared in 55 to 77% of 994 responses. That is roughly the ceiling, and those are globally famous consumer brands. Second, if you sell services, expect to be named far more often than you are linked, because ChatGPT cites a source in under 4% of professional services answers.
What to do when AI gets your brand wrong
Being described incorrectly is worse than being absent, because a wrong description travels.
Wrong description, wrong pricing, wrong services
Find the source first. Run the brand prompt on Perplexity and read the citation list. In most cases the error traces to one specific page: an old directory listing, a stale press release, an outdated partner page, or your own about page that you have not touched since 2022.
Then fix in this order: correct or remove the source page, update your own source-of-truth pages, refresh your entity records so the correct version outnumbers the incorrect one, and publish a page that states the correct fact plainly under a question-form heading. “How much does [brand] cost?” answered directly, in text, on a crawlable page.
Confused with a similarly named company
This is an entity problem and it needs machine-readable signals, not more copy.
Ship Organization schema with legalName, foundingDate, founder, full address and a complete sameAs array. Create or correct a Wikidata item. Make sure your founder’s LinkedIn, your Crunchbase entry and your website all point at each other. Where the collision is severe, use a consistent disambiguator in your own copy: “CrazyDi, the Search Everywhere Optimization agency based in Mumbai” rather than “CrazyDi.” Machines resolve entities from co-occurring attributes, so give them attributes only you have.
When AI recommends your competitor instead of you
You run prompt 1. Three competitors appear. You do not. The useful response is diagnostic rather than emotional, and it starts with one question: are they there because they are bigger, or because they are better documented?
Check what got cited when they were named. If it is their own comparison page, their pricing page, or a Reddit thread where a customer described a real result, that is documentation, and documentation is copyable. If it is Forbes, Gartner and three years of national press, that is authority, and you will not close that gap this quarter. In our experience most displacement is documentation: competitors get named because they wrote the page that answers the question while you wrote a page about your services.
So take the exact prompts where they win and write the pages that answer them, more specifically than they did. Where you are being compared unfavourably, the framing usually comes from a comparison page the competitor wrote or sponsored, which means you are being described in their words using their criteria. Write the honest version yourself: a page that names the competitor and states clearly where they are the better choice and where you are gets retrieved more often than a one-sided page, and reads as more credible.
Two more moves. Narrow the ground, because you may not win “best SEO agency” but “best SEO agency for B2B SaaS companies in India” has fewer defenders and a more qualified buyer. And get into the third-party sources that named them, since those roundups are retrieved as a unit and inclusion is often a matter of asking.
One thing worth knowing before you panic: engines agree on which brands belong in a category far more than they agree on which sources to quote. If a competitor beats you on every platform, that is a brand-level problem. If they beat you on one, it is a source-level problem on that platform, and much cheaper to fix.
The fix, in priority order
Sequence matters more than the individual tactics, and this is where most advice falls apart: it hands you thirty things with no order. Each tier below assumes the one above it is done. Effort is tagged Solo, Small team (a marketer plus a developer), or Agency-supported.
Week 1: the two things with the fastest payoff
1. Open the crawler gates. [Solo, 30 minutes]
Check that your robots.txt is not blocking the AI search crawlers: OAI-SearchBot, PerplexityBot, Claude-SearchBot, bingbot and Googlebot. Check your CDN and firewall too, because bot-blocking presets stop AI crawlers without touching robots.txt and most teams never look there.
This single fix takes a brand from a hard zero to a possible answer. If your citation rate came back at 0 to 5%, start here before writing anything.
2. Fix your source-of-truth pages. [Solo, 2 to 3 hours]
Write one sentence that says what you do, who you do it for, and where. Then make that exact sentence appear on your homepage, your about page, your LinkedIn Company Page, and the meta description of each. Correct your services list and, if you publish pricing, your pricing page.
AI systems assemble an answer from multiple sources. When those sources disagree about what you do, the model hedges, picks the loudest source, or leaves you out because a competitor’s description is cleaner. Giving it one clean version to find is the highest-return two hours in this guide.
Weeks 2 to 4: entity and structure work
Claim and correct your entity records. [Small team] Wikidata, Crunchbase, your Google Business Profile, and the review platforms your category uses. Same sentence everywhere.
Ship Organization schema with a complete sameAs array. [Small team] Link every profile you just cleaned up. This is how you tell a machine that the LinkedIn page, the Crunchbase entry and the website are one company.
Build the pages that answer comparison and problem prompts. [Small team] Every prompt on your sheet where you scored zero and a competitor scored one is a page you have not written.
Restructure existing pages for extraction. [Small team] Question-form headings, a direct answer in the first two sentences under each, tables for anything comparative, and a real FAQ block.
Month 2 onwards: the off-site work that compounds
Be present where the citations already come from. [Agency-supported] Your Perplexity source list told you exactly where. Answer real questions as a named person with a real title, and do not spam. Explainer videos targeting your problem prompts belong here too, because video is retrieved heavily in some categories.
Third-party listicles and digital PR. [Agency-supported] “Best [category] companies” roundups are retrieved directly and quoted almost verbatim. Getting into the ones that already rank is faster than writing your own.
Reviews with substance. [Small team] Volume matters less than review text that names your category, your use case and your customer type in the reviewer’s own words.
Not sure which tier you are actually on? CrazyDi’s free visibility audit runs this exact diagnostic across ChatGPT, Perplexity, Google AI Mode and Copilot, benchmarks you against your named competitors, and tells you where to start. Book Your Free Visibility Audit.
Frequently asked questions
How do I check if my brand is mentioned in ChatGPT?
Open ChatGPT in a logged-out or incognito window and run the five prompts above: three category and comparison questions to see whether you appear at all, and two brand questions to see whether what it says about you is accurate. Record whether you appear, which competitors appear, in what order, and whether a source link is shown. The logged-out session matters, because ChatGPT personalises answers from your history and a signed-in check will show you yourself.
How do I track brand mentions in ChatGPT over time?
Build a fixed set of 20 to 30 prompts drawn from your real sales conversations, run each three times per platform in fresh sessions, and log every run in a spreadsheet. Repeat monthly with the same prompts, because AI returns a different brand list more than 99% of the time and only repeated identical runs give you a comparable number.
Why is my brand not showing up in ChatGPT?
The common causes, in the order worth checking: your robots.txt or CDN blocks the AI search crawlers; your brand description is inconsistent across your website, LinkedIn and directory listings, so the system cannot resolve you into one entity; you have too few third-party mentions on the sources that platform cites; you have no pages answering the comparison and problem questions buyers ask; or your company is too new to have accumulated mentions. The first two are fixable in a week. The rest take months.
How long does it take to start appearing in AI answers?
Starting from near-zero visibility, expect 4 to 8 weeks for crawler and consistency fixes to register, 3 to 6 months for entity and content work to move your citation rate, and 6 to 12 months for off-site mentions to compound. Anyone promising results in two weeks is either fixing a blocked crawler, which is real and fast, or overselling.
Do I need a paid AI visibility tool?
No. Not to start, and possibly not at all. Everything here runs on a browser, a spreadsheet and about two hours a month, and it gives you citation rate, share of voice, sentiment, accuracy and average position: the same metric set the paid tools report, calculated from data you can inspect. Two free sources add to it. Google Search Console’s generative AI performance report shows impressions in AI Overviews and AI Mode, and Bing Webmaster Tools’ AI Performance report shows citation counts and the queries Copilot used to find you.
Paid tools become worth it when the manual run costs more than the subscription, usually once you are tracking more than 100 prompts or managing several brands. Buying one before you have run the manual audit means paying for a number you have no way to interpret.
Start with the number, not the tactics
You cannot improve what you have not measured, and almost nobody in this space is measuring. Run the five prompts today. Build the 25-prompt set this week. Log your first month and get a real baseline.
Then fix in order: crawler access and a consistent description first, entity and structure next, off-site mentions last.
The gap between brands that instrument this and brands that guess is currently enormous, and it is closing. Almost every competitor in your category is guessing.
