You rank on Google. You have solid pages. And yet you ask ChatGPT about your own category, and it names a competitor instead of you. If you have quietly wondered "why am I invisible in AI search," take a breath, because this is fixable, and the first job is not to fix anything. It is to find out which of four things is actually broken.
Here is the good news. AI invisibility almost always traces back to one of four layers: crawl access, retrieval eligibility, content structure, or authority signals. Most teams panic and start rewriting content, when the real problem is a blocked bot they could fix in five minutes. This guide is a triage. You will work through the four layers in order, run a quick pass or fail test at each one, and walk away knowing the single AI visibility root cause that is holding you back. It is the calm, ordered way to diagnose AI citation problem symptoms before you touch a single fix. Then you open the right fix, not all of them.
Let's find your layer.
Why AI search is not Google ranking
Before you diagnose anything, sit with one idea, because it changes everything that follows. AI engines do not rank pages. They select a small handful of sources to cite inside a written answer. That selection has to clear four separate gates, and missing any one gate drops you from the answer no matter how well you do on the others.
Google leans heavily on backlinks. AI engines lean more on brand mentions. In Ahrefs' study of 75,000 brands, web mentions of a brand correlated far more strongly with AI visibility than backlinks did. So a page sitting at position one on Google can be completely absent from the AI answer for that exact query. That gap is the reason "we rank everywhere but ChatGPT ignores us" is the most common complaint in AI search right now.
Feeling behind? You are not. Almost nobody has this fully figured out yet. You just need to know which gate is closed.
Step 1: Map the buyer prompts you should be winning
You cannot diagnose invisibility against "AI search" in the abstract. You diagnose it against specific questions your buyers actually type. So start by writing them down.
List 10 to 20 real prompts a buyer would ask an AI engine on the way to choosing a product like yours. Mix the obvious ones ("best CRM for SaaS") with the buying-stage ones ("how do I switch from X to Y"). These become your test set for every layer below.
How to tell it is done: you have a written list of prompts, and each one is phrased the way a person talks to ChatGPT, not the way you'd type a keyword into Google.
Where people go wrong: they test three vanity prompts with their brand name in them and declare victory. AI answers your branded queries easily. The money is in the unbranded, category-level prompts where buyers have not heard of you yet.
This is one place a platform earns its keep. DeepSmith's Discover Prompts generates a starter set of buyer prompts from your product, persona, and buyer-stage context, so you are not staring at a blank page trying to guess what people ask. You still own the list. The tool just gets you moving faster.
Step 2: Confirm AI bots can actually reach your pages
Now the fastest, cheapest check. Is my site crawlable by AI? If an AI bot cannot fetch your page, nothing else matters, because the page cannot be cited. Crawl access is binary. You are allowed, or you are not. There is no partial credit.
Pull up your robots.txt and your Cloudflare or WAF settings, then confirm the AI user-agents are allowed. The ones that matter for citations today are the search and live-fetch bots: OAI-SearchBot and ChatGPT-User for ChatGPT, PerplexityBot and Perplexity-User for Perplexity, Claude-SearchBot and Claude-User for Claude, and Google-Extended for Gemini grounding. Fetch a key page manually with one of those user-agent strings and confirm you get a clean 200, not a 403 or a timeout.
How to tell it is done: a manual fetch as OAI-SearchBot or PerplexityBot returns your real page content, and no rule in robots.txt or your firewall disallows those agents.
Where people go wrong: they block GPTBot to protect their content and assume they have blocked ChatGPT. They have not. GPTBot is the training crawler. OAI-SearchBot and ChatGPT-User are the ones that pull pages into answers. Blocking training does not block citation, and confusing the two is a common self-inflicted wound.
Common mistake: the silent Cloudflare block. In mid-2025 Cloudflare flipped its default so new domains block AI crawlers unless the owner explicitly allows them. Over a million sites had opted into blocking. If you are behind Cloudflare and nobody deliberately opened the gate, you may be turning away every AI crawler without knowing it. A clean robots.txt is not enough here. Check the Cloudflare dashboard's AI crawler setting directly.
If a bot is blocked, stop. You found it. Fix robots.txt and your Cloudflare rules, re-test with a manual fetch, and wait for a recrawl. That is the whole job for this layer. The question is my site crawlable by AI gets a clean yes here only when every citation bot passes a real fetch. If they all pass and you are still missing, move down to Step 3.
Step 3: Check whether the engine even considers your page
Say the bots can reach you, but the answer still leaves you out. Welcome to AI retrieval troubleshooting, the layer most people skip straight past. Retrieval is the engine's step of building a candidate set: the short list of pages it even considers for a given prompt. If your page never makes that list, no amount of on-page polish gets you cited.
Run each prompt from Step 1 in ChatGPT, Perplexity, and Gemini with browsing on. For every prompt, log three things: does the engine mention your brand, does it cite your page as a source, and what are the top three URLs it does cite. That last column is gold, because it shows you exactly which competitor pages are winning the candidate set you want into.
How to tell it is done: you have a simple grid, prompts down the side, engines across the top, with mentioned, cited, and top-three-URLs filled in for each cell.
Where people go wrong: they read "not cited" as "my content is bad" and start rewriting. But if the engine returns competitors and not you, the usual culprit is not your prose. It is retrieval. Three things commonly break it: your page uses different words than the prompt (the buyer asks for "CRM for SaaS," your page says "relationship software for subscription businesses"), your internal links do not connect the page to its topic cluster, or no third-party page mentions your brand alongside the category, so the engine has no reason to associate you with it.
Pro tip: the grid itself tells you where to go next. If the engine mentions your brand but does not cite your page, that is a structure problem, so head to Step 4. If it never mentions you at all, the issue is retrieval here or authority in Step 5.
Running this by hand across a dozen prompts and three engines, every week, is exactly the kind of repetitive work that quietly falls off the calendar. This is where the Prompts and Pages views in DeepSmith turn a manual audit into a scheduled job: it checks your prompts on a cadence and reports which of your pages actually earn citations, so the grid stays current without you rebuilding it each Monday.
Step 4: Test whether your answer is extractable
Passed the first two gates? Nice. Now we check extractability, which is a fancy word for a simple question: after the engine fetches your page, can it lift a clean, self-contained answer out of it? Engines do not summarize your whole article. They grab discrete claims. If your best answer is buried under a long windup, it gets passed over.
Try this spot-check on any page. Copy the first sentence after a heading into a blank doc. Does it stand on its own and answer the heading? If yes, that section is extractable. If it only makes sense after reading three paragraphs above it, that section fails.
How to tell it is done: the first 40 to 60 words after your H1 fully answer the page's core question, and every H2 contains at least one claim, a definition, a number, or a comparison, that can be lifted and still stand alone.
Where people go wrong: they front-load personality instead of the answer. Roughly 44% of ChatGPT citations come from the first third of the page, so a clever slow-burn intro is quietly costing you. Thematic headings hurt too. "How AI engines pick sources in 2026" is extractable. "The shift nobody saw coming" is not.
Common mistake: betting on FAQ schema to save weak content. Plenty of teams add FAQPage markup and wait for citations that never come. Large language models tokenize that JSON-LD as plain text and do not treat it as a private citation channel. Schema can help Google's AI Overviews, but it will not rescue a page whose answers are not extractable in the first place. Structure the content, then consider schema as plumbing, not a lever.
If you pass crawl and retrieval but fail this test, you found your layer. The fix is structural: front-load your answers, swap thematic headings for specific ones, attribute your stats inline, and add tables where you are comparing things. This is also where a production tool does real work. DeepSmith's Writer builds articles extractable by default, with the answer up top, a liftable claim under each heading, headings that name the query, internal links, and metadata, so extractability is baked in during writing instead of retrofitted after.
Step 5: Weigh your authority signals
Last gate, and the slowest one, so save it for last on purpose. Two pages can be equally extractable, and the engine will cite one and ignore the other. The tiebreaker is authority, and AI authority is not the same as Google authority. It is built from brand mentions across the web, E-E-A-T signals on the page, topical depth, and earned media.
Run five quick checks. Do you have more unlinked brand mentions on the web than the competitor winning your prompt? Does the page show real E-E-A-T (a named author with credentials, a last-updated date, sourced claims, a fast and secure page)? Do you have genuine topical depth on this subject, or one thin page? Are you present in earned media, the third-party lists, reviews, and forum threads where AI actually pulls most of its sources? And do you have any YouTube footprint at all?
How to tell it is done: for each buyer prompt, you can say whether your brand out-mentions the top competitor across the open web, and you can point to the E-E-A-T signals on the page.
Where people go wrong: they treat authority like an on-page checkbox. Most AI citations, across several independent studies, come from earned media rather than a brand's own site. Your own domain contributes a surprisingly small slice of the sources an engine uses to describe you. So if you have poured everything into your blog and nothing into being mentioned elsewhere, the engine has little outside evidence that you matter in this space.
Pro tip: authority is the one layer partly outside your control, and that is okay to say out loud. You can publish strong E-E-A-T pages this week. You cannot make a major publication mention you on demand. The fastest levers inside this layer are the ones you own: named authors and dates and sourced claims on-page, plus a real YouTube and community presence over time. The point of the triage is to know when authority is the cause, so you invest months of earned-media work only when it is genuinely the blocker.
Seeing competitors cited on prompts where you are absent, over and over? That is your signal. DeepSmith's AI Visibility Competitors view tracks which competitor pages earn citations, on which prompts, so you can see who the engine trusts and reverse-engineer why, instead of guessing.
Step 6: Read the triage result and open the right fix
You have run all four gates, which means you can finally diagnose AI citation problem cases by evidence, not by guesswork. Now read the result, because your answer maps cleanly to one next move. Do not open all four fixes at once. Open the one your triage pointed to.
- A bot was blocked or the page was unreachable. Open the technical fix. Correct robots.txt, your Cloudflare setting, and any firewall rule, then re-test with a manual fetch. Minutes of work, biggest payoff.
- Crawl passes, but the brand or topic is invisible to the engine. Open the retrieval fix, the heart of AI retrieval troubleshooting. Align your on-page language with how buyers phrase the prompt, repair your internal link graph, and start earning third-party co-mentions with the topic.
- You are retrieved but never cited. Open the structure fix. Front-load answers, make headings specific, attribute stats inline, and add tables for comparisons.
- Competitors get cited on your prompts and you do not. Open the authority fix. Build unlinked mentions, add E-E-A-T signals, deepen your topical coverage, and invest in earned media and YouTube.
- Sometimes cited, sometimes not. Run the triage per prompt. You are probably passing on some queries and failing on others, and the fix is a blend across layers.
Notice the order once more, because it saves you real money and time. Crawl fixes take minutes. Structure fixes take days. Authority fixes take months. If you diagnose in that sequence, you never spend a quarter on an earned-media program to solve what turned out to be a single blocked user-agent.
What to do next
You do not need to boil the ocean. The question you started with, why am I invisible in AI search, now has a real answer path. You know your layer, and you know the AI visibility root cause behind it. Pick your most important buyer prompt, run it through the four gates this week, and write down where it fails first. That one finding tells you the only fix worth starting.
If keeping this diagnosis current across dozens of prompts and five engines sounds like a job you will never get to, that is exactly the gap DeepSmith was built to close. It tracks your mention and citation rates across ChatGPT, Perplexity, Gemini, Claude, and Google AI Mode, shows you where you are missing, and produces publish-ready, extractable content to fix the structural gaps, all from the same data. Want to see your own layer, with real numbers? Start a free trial and run your first prompts today.
