If you rely on organic traffic, the GPT-5 ChatGPT search change raises a clear question. Does the new model change who gets found and cited in ChatGPT answers? The short answer: GPT-5 changes the model that reads and writes, not the rules for getting into ChatGPT search. Your site still needs to be crawlable by OpenAI’s search crawler. Your pages still need to give clear answers that can be checked and are worth citing.
OpenAI released GPT-5 on 7 August 2025 and made it the default model in ChatGPT. It rolled out to Free, Plus, Pro and Team users that day. Below I cover what OpenAI actually claimed, what is known about how ChatGPT search finds and cites websites, and what to check on your own site this week.
What OpenAI Announced on 7 August
For most people, GPT-5 replaced the old model picker. Signed-in users used to choose between GPT-4o, o3, o4-mini and others. Now they get GPT-5 by default, and OpenAI says it has replaced those older models.
The main points from OpenAI’s announcement:
- One system, two modes. GPT-5 combines a fast model for everyday questions with a deeper reasoning model for harder problems. OpenAI calls the second one GPT-5 thinking.
- A real-time router. OpenAI describes a router that picks which mode to use. It looks at the type of conversation, how complex it is, the tools needed and what the user asks for, such as “think hard about this”.
- Access for everyone, with limits. Free users get GPT-5, then move to a smaller GPT-5 mini model once they hit their usage limit. Plus users get higher limits. Pro users get unlimited GPT-5 plus a GPT-5 Pro option. Enterprise and Edu access was due the following week.
- Thinking when you want it. Paid users can pick GPT-5 Thinking from the model picker. OpenAI also said full reasoning for free users might take a few days to roll out.
For search marketers, the important part is scale. Every ChatGPT user now gets the same model family. That includes free users, who are the largest group. Whatever GPT-5 does when it searches and cites sources is now the default across the product.
OpenAI’s Accuracy and Hallucination Claims
OpenAI put a lot of weight on factual accuracy in this launch. A hallucination is when an AI states something false as if it were true. The figures below are OpenAI’s own, from its own tests. They have not been checked by anyone independent, so I treat them as the company’s claims.
- With web search turned on, OpenAI says GPT-5’s answers are about 45% less likely to contain a factual error than GPT-4o’s.
- When thinking, again with web search on, OpenAI says GPT-5’s answers are about 80% less likely to contain a factual error than o3’s.
- On open-ended factuality benchmarks (LongFact and FActScore), OpenAI reports that GPT-5 thinking produced about six times fewer hallucinations than o3.
OpenAI also reported less deception. On conversations it describes as typical of real use, it says deceptive reasoning answers fell from 4.8% for o3 to 2.1% for GPT-5. It also says GPT-5 is better at admitting when it cannot finish a task, instead of confidently making something up.
A third claim matters for content. OpenAI says GPT-5 is less sycophantic, meaning less likely to just flatter or agree with the user. In its targeted tests, sycophantic replies dropped from 14.5% to under 6%. In plain terms, the model is meant to agree with the user less by default and push back more when the facts do not support them.
Why the “With Web Search” Detail Matters
Look at how OpenAI measured the headline accuracy figures. They were taken with web search turned on. In other words, OpenAI presents GPT-5’s accuracy partly as the result of basing answers on web pages it finds.
I think that signal is worth taking seriously. If the model is expected to back up its answers with web sources, then the pages it finds are doing real work. Clear, accurate, well-sourced pages are what that kind of system needs most.
OpenAI’s announcement did not describe any change to how ChatGPT search ranks or picks sources. So be careful of anyone claiming GPT-5 brought in a new “ranking algorithm” for ChatGPT search. OpenAI has not said that.
How ChatGPT Search Finds and Cites Websites
ChatGPT search has existed since October 2024. OpenAI opened it to all logged-in users in December 2024. In February 2025 it made it available to everyone where ChatGPT is offered, with no sign-up needed. GPT-5 now sits on top of that same search experience.
Here is what OpenAI has documented about how it works.
When ChatGPT Decides to Search
According to OpenAI’s ChatGPT search announcement, ChatGPT chooses to search the web based on what the user asks. Users can also force a search by clicking the web search icon.
So not every prompt starts a search. Questions about current events, prices, local businesses, recent products or anything time-sensitive are the most likely to send ChatGPT to the web. A question about a timeless idea may be answered from the model’s own knowledge, with no citations at all.
Where the Results Come From
OpenAI says ChatGPT search draws on third-party search providers, plus content that partners send directly. It named news partners such as the Associated Press, Reuters, the Financial Times and others.
The practical lesson is that being visible in traditional search indexes still helps. OpenAI has not published which providers it uses, so I won’t guess. But a page that cannot be found in mainstream search engines has fewer ways into ChatGPT’s results.
How Citations Appear
When ChatGPT searches, answers include links to the sources it used. Users can click a Sources button under the answer to open a sidebar listing every reference. These citations are your traffic opportunity. They are the clickable route from a ChatGPT answer back to your site.
The Three OpenAI User Agents You Should Know
OpenAI documents its crawlers in its overview of OpenAI crawlers. A user agent is the name a crawler uses when it visits your site. Each one is controlled separately in robots.txt, and mixing them up is the most common mistake I see.
| User agent | What it does | What blocking it means |
|---|---|---|
OAI-SearchBot |
Crawls sites so they can be shown in ChatGPT search results. | Your pages won’t be shown in ChatGPT search answers, though they may still appear as navigational links. |
GPTBot |
Crawls content that may be used to train OpenAI’s foundation models. | Signals that your content should not be used for training. It does not remove you from search. |
ChatGPT-User |
Visits pages when a user’s action in ChatGPT or a custom GPT asks for them. | Not used for automatic crawling or to decide search eligibility. OpenAI notes robots.txt rules may not apply. |
The key point: search and training are separate choices. You can allow OAI-SearchBot so you show in ChatGPT search, while blocking GPTBot so your content isn’t used for training. OpenAI says it can take about 24 hours for its search systems to reflect a robots.txt change.
If you want a closer look at visits that users start, I covered how ChatGPT-User and agent browsing behave in what ChatGPT Agent means when AI browses your website.
What GPT-5 Could Change for ChatGPT Search Visibility
This section is my own reading as someone who does this work, not something OpenAI has said. I’ve kept it to what follows logically from OpenAI’s own claims.
More Reasoning Can Mean More Searching
GPT-5’s router decides when to think harder. OpenAI says it weighs “tool needs” when it does so, and web search is one of those tools. If more complex questions go to the thinking mode, it is fair to expect those answers to lean on web sources. That is even more likely since OpenAI frames its accuracy gains around web search being on.
For publishers, that favours pages that answer detailed, multi-part questions properly. A thin page that repeats a definition is unlikely to be the source a reasoning model needs to finish a complex answer.
Pressure for Accuracy Favours Content You Can Check
A model trained to make fewer factual errors and to admit when it is unsure has a reason to prefer sources that state facts plainly. Pages that bury the answer, hedge everything or mix opinion with fact make that harder.
Practical examples of content that can be checked:
- A pricing page that lists real prices, what’s included and the date they were last updated.
- A product page with specs in a table, not scattered through marketing copy.
- A how-to guide with numbered steps and the exact menu names or settings a reader will see.
- A local business page with the address, opening hours and service area written out in text.
Less Sycophancy Changes “Best of” Content
If GPT-5 really is less inclined to tell users what they want to hear, then pages built only on sales claims have less to offer it. “The best plumber in the city” is not a fact a model can check. “Gas Safe registered, 24-hour call-out, covering these postcodes” is.
This is the same direction Google has pushed for years with helpful content. In my view, good SEO copy and good AI-citation copy are becoming the same thing.
What Has Not Changed
- OpenAI has not announced a new way to submit pages to ChatGPT search.
- No paid placement in ChatGPT search results was announced as part of GPT-5.
- Robots.txt still controls access. Blocking
OAI-SearchBotstill keeps you out of search answers. - No one outside OpenAI knows exactly how it picks which sources to cite. OpenAI has not published a ranking guide.
GPT-5 ChatGPT Search: What Site Owners Should Do Now
None of this needs a rebuild. It needs a check of a few technical basics, and tighter content on the pages you most want cited.
1. Check robots.txt for OpenAI’s Crawlers
Open yourdomain.com/robots.txt. Look for any rule that names OAI-SearchBot, or a blanket User-agent: * block that would catch it. If you want to show in ChatGPT search but keep content out of model training, a setup like this does both:
User-agent: OAI-SearchBot
Allow: /
User-agent: GPTBot
Disallow: /Only block GPTBot if that is a deliberate business choice. Plenty of sites are happy to allow both.
2. Check Your Firewall and CDN, Not Just robots.txt
Some security tools and CDN bot settings block AI crawlers by default. OpenAI publishes IP ranges for each crawler and recommends allowing OAI-SearchBot requests from those ranges. If your server logs show no visits from OAI-SearchBot at all, check if something upstream is rejecting it. This is the first place I look when a site seems invisible to ChatGPT.
What your server sends back matters too. A crawler that keeps hitting errors or redirect chains will struggle to fetch your pages. My guide to HTTP status codes and SEO explains what each response tells a crawler.
3. Make Sure Key Content Is in the HTML
If your prices, specs or answers only show after JavaScript runs, or sit behind tabs that load on click, crawlers may not see them. Check the raw HTML of your most important pages. If the core answer isn’t there, fix that first.
4. Lead With the Answer
Put a direct, factual answer near the top of each page that targets a question. Follow it with the detail, the steps and the proof. This helps human readers, Google’s featured snippets and AI systems for the same reason. The answer is easy to find and easy to quote.
5. Make Facts Specific and Dated
Add “last updated” dates to pages where freshness matters. Replace vague claims with figures you can stand behind. Cite your own primary sources where you have them. A model aiming for fewer factual errors has more reason to cite a page that shows its working.
6. Track ChatGPT Referral Traffic
ChatGPT referrals show in analytics as traffic from chatgpt.com. Links from ChatGPT often carry a utm_source=chatgpt.com parameter. In GA4, filter session source for chatgpt.com and compare the weeks before and after 7 August. If you need a starting point, my walkthrough on building organic search reports in GA4 shows how to set up an exploration you can adapt for AI referrals.
Expect the numbers to be small for most sites, and noisy for the first few weeks while the rollout settles. Look for direction, not exact attribution.
7. Test the Questions Your Customers Ask
Ask ChatGPT the questions your customers really ask, with search turned on, and open the Sources sidebar. Note which sites are cited, what format their pages use and what facts they state that yours do not. In my opinion, this is the most useful competitor research you can do right now, and it costs nothing.
Keep in mind that answers change between sessions and users. Run each question several times. Look for patterns instead of drawing conclusions from one result.
Common Mistakes to Avoid
- Blocking all AI bots with one rule. A blanket block that catches
OAI-SearchBotremoves you from ChatGPT search answers. That may not be what you meant. - Thinking that blocking GPTBot hides you from ChatGPT search. It doesn’t. Search eligibility is controlled by
OAI-SearchBot. - Chasing “GPT-5 optimisation” tricks. OpenAI has not published ranking factors. Anyone selling a GPT-5 hack is guessing.
- Ignoring Google. ChatGPT search relies partly on third-party search providers. So strong traditional SEO helps your visibility in both.
- Judging the impact too early. The rollout is days old, free-tier reasoning was still rolling out, and usage patterns will shift. I’d give it a few weeks before changing strategy.
Conclusion
GPT-5 ChatGPT search is the same search product, with a stronger model that cares more about accuracy reading the results. OpenAI’s claims of fewer factual errors with web search on point one way. Well-structured pages with facts you can check are what the system needs to cite.
Your next step is simple. Check that OAI-SearchBot can reach your site. Then review your top pages for direct answers and specific facts. If your site already earns rich results in Google, the same habits apply. My guide to SERP features and how to earn them is a good place to tighten that up.
