Key takeaways
- AI engines are not judging your brand. They summarize what their sources say, so the repair work happens on those sources, not in the chat window.
- A 90-day plan has four phases: audit (days 1-14), fix your own pages (days 15-45), fix third-party sources (days 30-75), and re-measure with identical parameters (days 76-90).
- Landing pages, docs and comparisons stay cited longest in ChatGPT, while news fixes fade in about five weeks. Put your corrections where they last.
- A word like "limited" often signals a gap in what AI can find about you, not a real product flaw. Fill the gap before you rewrite anything.
- Changes are not instant. Engines have to re-crawl, and training-based impressions can lag even longer.
Why arguing with ChatGPT does nothing
You type your brand name into ChatGPT and the description sounds like someone else's company. Maybe it says you are "limited," or expensive, or it still lists a product you retired two years ago. The instinct is to correct the chatbot. Don't bother. A correction in one chat does not change what it says to the next person.
What works is treating AI perception as a downstream symptom. The model builds its paragraph about you from training data and, for engines like Perplexity, ChatGPT Search and Google AI Overviews, from live retrieval. Change the pages those systems read and cite, and the paragraph changes on later crawls. One 2026 guide from Kairosy puts it well: the improvement work is mostly upstream, and perception follows once the sources are fixed.
The flip side is that this takes patience. Changing a source does not change an answer the same afternoon. Plan for weeks, not days, which is why a structured 90-day plan beats ad hoc fixes.
What AI brand perception actually is
AI brand perception is how assistants describe and characterize your brand at the moment someone is deciding. It is measured across many prompts, not one answer, because responses vary by engine, by country, and by model update.
Searchable analyzed 1,943,365 negative descriptor phrases across 6,783 brands (July 5 to August 18, 2026). The most repeated negative word was "limited", which showed up for 898 brands, more than any price or service complaint. Searchable's reading is that this usually points to a gap in what AI engines can find about a brand rather than a real product weakness. They also found that branded prompts surface "limited" first, while unbranded category prompts surface pricing. Across 11,289 brands with at least 100 observations, the median sentiment score was 66.9 and negative descriptors were only 1.6% of descriptors for the median brand.
Two things follow. Most negative perception is fixable, and you have to separate branded prompts ("what is Brand X like?") from unbranded ones ("best tools for Y"), because the problems differ.

The four things worth measuring
- Visibility: do you show up at all for relevant prompts? If rarely, fix visibility before sentiment.
- Sentiment: the tone of the language used about you.
- Descriptors: the specific adjectives and claims repeated across engines.
- Accuracy: whether the facts (pricing, features, locations, product lines) are right.
Read together, they tell you where the problem is. Low visibility is a coverage problem. Negative descriptors are a reputation or content problem. Wrong facts are a source-consistency problem.
Days 1-14: audit what AI says today
The first two weeks are about building a baseline you can compare against on day 90. The Search Engine Land 90-day plan for off-page AI visibility makes the same point: rerun the audit with the exact parameters you set at the start (same prompts, engines, competitor set, and window length). If you change the parameters, you cannot tell whether anything improved.
Step 1: write your prompt set
Build 30 to 60 prompts split into three groups:
- Branded: "What is [brand]?", "Is [brand] good for small teams?", "[brand] pricing", "[brand] reviews".
- Competitor comparison: "[brand] vs [competitor]", "alternatives to [competitor]".
- Unbranded category: "best [category] for [use case]" in the phrasing your buyers actually use.
Run them on ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews at minimum. Use the same country and language you sell in. Manual spot checks work for a first look, but a one-time search tells you what one engine said once. It will not surface patterns or catch errors across product lines.
Step 2: extract the claims
For each response, record every individual claim about your brand: "affordable for small teams," "mobile app lacks features," "founded in 2012," "no free plan." Group repeated claims and note which engines say each one. You now have a list of descriptors with frequencies. Mark each as accurate and positive, accurate and negative, or inaccurate.
Step 3: trace each claim to a source
For every negative or wrong claim, find where the engine got it. Look at the citations in Perplexity, ChatGPT Search and AI Overviews. Common culprits are:
- An old pricing page or an outdated comparison article.
- A review site profile you never claimed.
- A forum thread or Reddit post with a loud complaint.
- A competitor's comparison page.
- Your own site, saying something vague or contradicting itself.
That last one is more common than people expect. The "limited" finding suggests many brands simply have thin product, pricing and feature pages for engines to read.
Tools for the audit
You can do this in a spreadsheet, but tooling saves real time once you pass a few dozen prompts. Promptwatch collects responses from the real interfaces of ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews by country and language, and its Brand Narrative feature pulls the individual claims AI makes about you out of every tracked response and groups them by theme like pricing, features and reputation. Fact Check, on higher plans, lets you mark a claim inaccurate and then finds the pages on your site that state it.

Other options in the same space include Semrush, whose AI Visibility Toolkit includes a Perception view with key sentiment drivers, and Searchable, which splits descriptors by branded and unbranded prompts.

Waikay is also built specifically around brand perception tracking.
| Tool | Strongest angle for perception work | Best for |
|---|---|---|
| Promptwatch | Claim-level Brand Narrative, Fact Check, then content and tracking of the fix | Teams that want to audit and repair in one product |
| Semrush AI Visibility Toolkit | Sentiment drivers inside an existing SEO suite | Teams already on Semrush |
| Searchable | Descriptor tracking split by branded and unbranded prompts | Descriptor-level analysis |
| Waikay | Perception-focused tracking | Smaller perception audits |
| Spreadsheet and manual prompts | Free, slow | A first baseline under 30 prompts |
For wider comparisons, the GEO software directory at bestgeosoftware.com lists far more options.
Step 4: score and prioritize
Rank problems by two factors: how many engines repeat the claim, and how close it sits to a buying decision. A wrong price in a branded query matters more than a quirky adjective in a long-tail prompt. Pick the top five to eight issues. Trying to fix thirty at once is how plans stall.
Days 15-45: fix the sources you control
Your own site is the fastest lever. Promptwatch Data shows what engines read when someone asks about a company by name. In its report on which content types ChatGPT cites the most (Sep 28 to Oct 4, 2026), product pages made up 23.3% and landing pages 22.8% of citations for branded prompts, almost half combined. Reviews rose to 5.6%, more than three times their share across all prompts, and FAQ pages reached 4.8%.
So if AI describes you badly, your homepage, product pages and docs are the first place to look.
Make the facts easy to quote
- State pricing, plans, and who the product is for in plain language on product and pricing pages. Vague copy invites vague, hedged descriptions.
- Add a plain-language FAQ page that answers the exact questions in your prompt set, including the uncomfortable ones ("Does it have a free plan?", "How does it compare to X?").
- Retire or redirect outdated pages. If a 2022 pricing page still ranks, engines will read it.
- Keep specs and prices current, since product pages are the most cited type.
- Build honest comparison pages. Comparisons held at 72% still cited after 30 days in Promptwatch's half-life data, so they last.
Check that AI can actually read your site
A surprising share of perception problems are access problems. If AI crawlers get blocked or hit errors, the engine fills the blank with whatever third parties say. Check robots.txt, rendering, and server errors for bots like GPTBot, ClaudeBot and PerplexityBot. Promptwatch offers a free Agentic Readiness Check and AI crawler logs that show which pages bots read and where they fail. See Promptwatch Data's AI crawler traffic report for how crawler activity is shifting.
Where to publish for longevity
This matters for a 90-day plan. Promptwatch's content half-life report tracked every page ChatGPT cited on July 6, 2026 for 90 days. After 30 days, ChatGPT still cited 78% of landing pages and 76% of docs; after 60 days, 69% and 67%. Product pages, reviews and how-tos sat at about 55% after 60 days. News decayed fastest: about half of news pages were gone after 36 days.
My inference from that data: corrections made on landing pages, docs, and comparisons should persist through the plan, while a press-release fix may fade before day 90. Use news for short bursts and put durable corrections on durable page types.
Days 30-75: fix the sources you do not control
Once your own pages are clean, move outward. This phase is slower and less predictable, but it is where many negative descriptors actually originate.
Review and directory profiles
Claim and update every profile on the sites your buyers use in your category (G2, Capterra, Trustpilot, app stores, industry directories). Make sure descriptions, pricing and feature lists match your site. Then collect fresh reviews, because a handful of outdated one-star reviews can anchor an entire summary. Reviews are over three times more prominent in branded citations than in ChatGPT citations overall, so this effort pays off precisely where perception forms.
Third-party articles and listicles
If a popular roundup describes you with stale facts, contact the author with updated information. Listicles are 13.6% of all ChatGPT citations, so they carry weight. Be factual and brief, and offer a correction rather than a complaint. Digital PR can help here, and tools like Respona exist for outreach at volume.
Social and community sources: be realistic
Do not assume Reddit or LinkedIn will fix ChatGPT. In Promptwatch Data's Social Media Citations in AI Search, November 2026 (Oct 11 to Nov 9, 2026), social platforms were 0.64% of ChatGPT citations, 1.94% of Perplexity's and 13.16% of AI Overviews'. YouTube alone was 6.71% of AI Overviews citations. The report's own advice: don't rely on social posts to reach ChatGPT; prioritize YouTube for Google and Reddit and LinkedIn for Perplexity. Note that the window changed from 7 to 30 days versus the October report, so comparisons are not like-for-like.
For context on Reddit specifically, Promptwatch's data shows reddit.com's share of ChatGPT citations fell from about 4% to 0.5% on August 14, 2026. A Reddit thread complaining about you matters less in ChatGPT than it did, but it can still matter in Perplexity and AI Overviews.
Handle hallucinations separately
Some errors have no source. The model simply invented a feature or a founding date. These come from training data gaps. The remedy is the same, only slower: publish clear, consistent, crawlable facts in several places so retrieval has something accurate to cite, and wait for model refreshes. Where engines offer a feedback or report button, use it, though results are inconsistent and I did not find official guidance with commitments on turnaround.
Days 46-75: ship new content against the gaps
The "limited" finding deserves action. If engines call you limited, they probably cannot find enough about what you do. Look at the unbranded prompts where competitors appear and you do not. Each is a content gap: a question buyers ask that you never answered.
Publish pages that answer those prompts directly. Good candidates:
- Use-case pages for each audience you serve.
- Comparison pages with fair, specific claims.
- Documentation and how-tos, which hold up well (7.5% of ChatGPT citations are documentation, 8.7% how-tos).
- A glossary or explainer for category terms where you want to be the cited source.
Tools differ in how far they go. Monitoring-only products show gaps and leave the writing to you. Promptwatch's Content Agent plans, writes and publishes articles against your gaps to WordPress, Webflow or Framer (Business plan and up), then tracks whether those pages get crawled and cited. Content tools like Surfer SEO and Clearscope help optimize a draft but have no view of what AI cites.


| Approach | Shows the gap | Writes the fix | Checks if AI cited it |
|---|---|---|---|
| Manual prompts + spreadsheet | Slowly | You | Manually |
| Monitoring-only trackers (Otterly.AI, Peec AI) | Yes | No | Partly |
| Content optimization tools (Surfer, Clearscope) | No | Assists | No |
| End-to-end platform (Promptwatch) | Yes | Yes (Business+) | Yes |

If you want a hands-off option, an agency can run this for you. 1001 SEO Media handles AI search and GEO work and uses Promptwatch for it, but you can run the whole plan in-house with the steps here.
Days 76-90: re-measure and decide
Rerun the exact prompt set from days 1-14. Same engines, same countries, same competitor set, same window length. Then compare:
- Visibility and share of voice by engine, branded and unbranded separately.
- The descriptor mix: are engines using different adjectives, and do comparisons against competitors read more favorably?
- Accuracy: how many of your flagged wrong claims are gone?
- AI referral traffic and conversions, if you have analytics set up to separate them.
A score change by itself will not tell you which action caused it. So check the specific claims you targeted. If a negative phrase you attacked has dropped or disappeared, your fix worked and you can repeat the method on the next phrase. If nothing moved, check whether the source you changed has actually been re-crawled and cited yet.
One caveat on traffic attribution: ChatGPT stopped sending a referrer on October 8, 2026, so unattributed visits may hide AI traffic. Tagging links with utm_source=chatgpt.com and counting those visits closes part of the gap. Promptwatch's guide to attributing AI traffic covers the setup.
A condensed 90-day checklist
| Window | Focus | Output |
|---|---|---|
| Days 1-14 | Audit | Prompt set, claim list, source map, baseline scores |
| Days 15-30 | Own pages | Updated product, pricing and FAQ pages, crawl access fixed |
| Days 30-60 | Third parties | Updated review profiles, corrected roundups, outreach sent |
| Days 46-75 | New content | Pages answering gap prompts, comparisons, docs |
| Days 76-90 | Re-measure | Same-parameter rerun, before and after report |
Common mistakes
- Fixing everything at once. Pick five to eight claims and finish them.
- Judging by one engine. ChatGPT, AI Overviews and Perplexity cite very different sources.
- Rewriting copy when the real issue is that bots cannot read the page.
- Relying on news coverage for permanent fixes.
- Changing prompts between baseline and re-test, which makes the comparison meaningless.
- Expecting overnight movement. Retrieval-based engines move in weeks; training-based impressions can take longer.
What to do after day 90
Perception is not fixed once. Models update, competitors publish, and old pages resurface. Keep a monthly rerun of your core prompts and review the descriptor mix. Treat any new negative phrase as a ticket: find the source, fix it, confirm the change.
The free AI visibility report from Promptwatch is a quick way to get a first baseline before committing to the full plan. If you want to compare more options, ai-rank-tools.com lists AI rank and visibility trackers. Then block out week one, write your prompts, and start the audit.


