GPT-5: The Real Guide to Getting the Most Out of It
August 15, 2025
•
5
min read
So GPT-5 is here, and as usual, the internet has split into two camps: people who love it, and people who think it’s the worst thing to happen since autocorrect decided to “fix” your texts. Honestly, I think a lot of the hate comes from people not using it correctly. I’ve just spent seven days reading, watching, and pushing GPT-5 to its limits, so here’s the straight-to-the-point guide you actually need.
1. Unlock All the Hidden Models
When you open the model selector, you’ll probably only see:
Auto
Fast
Thinking
Pro (if you’re on the $200/month plan)
If you scroll down to "Legacy models," you might spot GPT-4o. But here’s the thing: there’s way more available.
Here’s how to unlock them:
Go to Settings
Open the General tab
Toggle Show additional models
Now, under GPT-5, you’ll see Thinking Mini, which is fantastic for copywriting and lighter creative work.
Which one should you use?
Auto: GPT will pick based on complexity. Simple = Fast, complex = Thinking.
Thinking: Better for complex, multi-step reasoning.
Fast: Cheap, efficient, and surprisingly good. Great default.
Thinking Mini: Like Thinking, but lighter and faster.
OpenAI’s routing system tends to pick the cheaper (Fast) option if possible. You can always start a chat with Thinking for context, then switch to Fast to save time and credits.
And yes, if you’re still emotionally attached to old favorites like GPT-4.1 or GPT-3.5, they live in the Legacy models section. OpenAI brought them back after a very vocal backlash.
2. Play With the New Personalities
GPT-5 now lets you pick a built-in personality. At first, I thought these were a bit of a gimmick, but the more I use them, the more I realize they’re surprisingly useful:
Cynic: Snarky, blunt, and great for no-nonsense feedback. My personal favorite.
Robot: Zero small talk. Just gets to the point. Perfect for coding or rapid drafting.
Listener: Supportive and empathetic. Weird to say about an AI, but some people genuinely use it for emotional support.
Nerd: Excited, thorough, and detail-obsessed. Perfect for learning something in depth.
Switch them up and see which fits your workflow. I’ve found Cynic is great for constructive criticism, and Robot is a huge time saver.
You can access all these personalities by going to the Customize ChatGPT.
3. Use the Canvas Feature for Apps and Dashboards
Canvas is GPT-5’s built-in interactive workspace. You can:
Build games
Create data dashboards
Make interactive reports
You can share these with anyone, even if they don’t have a ChatGPT account. This makes it a goldmine for lead generation, imagine sending an interactive report that doubles as a pitch.
4. (Gimmick Alert) Change Your Highlight Colour
It’s cosmetic, not functional, but you can change your GPT interface accent colour to blue, green, yellow, pink, orange, or purple. It’s the kind of feature Apple would hype up as a major release, even though it won’t make GPT-5 any smarter. Still, it does make the workspace feel less sterile.
5. Use the Prompt Enhancer to Fix Bad Prompts
If GPT-5 isn’t giving you what you want, stop blaming the model. Prompting is just communication, and bad prompts = bad results. OpenAI’s Prompt Enhancer (available in the Playground) takes your vague or poorly written prompt and turns it into something far more specific.
Poorly defined prompts are one of the biggest reasons people fail to get good AI outputs. In fact, research shows that improving prompt clarity can boost task accuracy by up to 30% (Stanford HAI). Use the enhancer and you’ll instantly see better results.
Bottom line: GPT-5 is incredibly capable, but only if you unlock all the tools and actually use them. Stop treating it like a magic trick and start treating it like the power tool it is.
Four Claude Code systems run my entire SEO workflow under one roof: keyword research, a content writer, a site health audit, and a content refresher. They feed each other like an SEO team would, ship blog posts that follow the capsule content method, and run on a Monday 9am cron. One repo, four prompts, free organic traffic.
30 sec readSkip to full article below
Why bother building an SEO system inside Claude Code?
Because most SEO tools force you to bounce between five tabs to ship one blog post, and the boring tasks (the ones that actually move rankings) are the ones you skip. I built four systems that connect under one roof and feed each other like an actual SEO team would. The result on one site was 14.4M impressions and 90,000 organic clicks. Two newer sites started pulling organic traffic the week they launched.
No ads, no agency, no five-tool stack. Just Claude Code running the work.
The four systems are keyword research, a content writer, an on-site audit, and a content refresher. Each one is a skill Claude Code can run on command, and each one outputs into a shared dashboard so the next system knows what the previous one did.
System 1 — Keyword research: Builds a keyword bank + fan-out cluster + content queue. Run once a month.
System 2 — Content writer: Drafts ranking-ready blog posts using the capsule method. Run weekly.
System 3 — On-site audit: Pulls a full health report via DataForSEO. Run fortnightly.
System 4 — Content refresher: Flags decaying or de-indexed posts to rewrite. Run monthly.
The trick is they share state. The keyword researcher knows what's already been covered, so the content writer can't cannibalise itself. The audit knows which pages exist. The refresher knows which ones are dying.
What do you need before you start?
You need a project folder, Claude Code, and a DataForSEO key. That's it.
A project folder on your local machine with your business info inside (services, locations, brand voice, USPs)
Hand Claude Code the repo link and tell it to install the systems for this business. Eight to ten minutes later, all four skills are wired up and pulling your business context. Auto-mode helps here so it stops asking permission every 30 seconds.
If you've never connected Claude Code to MCP servers before, watch my SEO Command Center setup video first, then come back.
How does System 1 (keyword research) work?
You type a service or topic, and Claude Code runs the keyword research skill, builds a fan-out cluster, and saves it to a dashboard.
In my example I ran it for “therapeutic gardening”. A couple of minutes later I had 31 keywords, a CSV file, and a live HTML dashboard. The dashboard has two things that matter:
A keyword bank of every keyword in the cluster, with status flags so Claude knows which ones it's already targeted (this is what stops content cannibalisation)
A fan-out cluster of supporting keywords that should appear as H2s or H3s inside the eventual blog post
So by the time System 1 is done, System 2 already knows exactly what to write and which headings to use. You only need to run keyword research once a month, or whenever you run out of content.
How does System 2 (the content writer) write ranking blog posts?
You tell Claude Code “write the next blog post”. It pulls from the keyword bank, drafts a post using the capsule content method, and outputs an MD file or publishes directly to your site.
Specifically, the content writer does six things automatically:
Injects your experience from the business files (first-person stories, real numbers, anything that smells like E-E-A-T)
Targets the primary keyword in the title and an H1, and the fan-out keywords in H2s and H3s
Writes ~70% in the capsule method (H2s phrased as questions, answered in the first one or two sentences)
Cites high-trust sources like government domains, official health bodies, primary research
Internally links across the site because it reads your sitemap
If your site is built on Astro, the post publishes itself to the live site without you ever opening a CMS. If you're on WordPress, you get a clean MD file to paste in, and the WordPress REST API can automate that part too.
Community win: Inside the AI Ranking community, Steven used a version of this exact workflow to index more than 800 local service pages, which generated 105 booked appointments in a single month from organic traffic alone. No ads.
Don't worry about making posts longer. Worry about making them better. The content writer was tuned for citation-readiness, not word count.
How does System 3 (the on-site audit) work?
You run “audit the site”, Claude Code calls DataForSEO, and you get a full health report inside the dashboard with prioritised fixes.
On the test site, it returned an on-page score of 97/100 and an SEO score of 99/100, plus a list of broken links and slow pages to fix. Total DataForSEO cost: about 48 cents. With the $5 free credit, you can run this 10 times before paying anything.
The audit also tells you exactly which fixes to do first. If your site is on Astro and Claude Code can edit it directly, you can tell Claude to fix them for you. If you're on WordPress, you do the fixes manually but at least you know what to fix.
Run this once a fortnight, definitely once a month. Most people skip on-page audits because the data is overwhelming. Claude's job is to do the distilling for you.
How does System 4 (the content refresher) work and why does nobody run it?
The content refresher reads your Google Search Console data, finds blog posts that are decaying or de-indexed, and tells you exactly which ones to rewrite. Almost nobody runs this, and it's the highest-ROI system of the four.
Here's why it matters: only around 60% of the blog posts you publish stay indexed. Google has been getting much stricter about what it keeps in its index, and “crawled, currently not indexed” is Google's passive-aggressive way of saying it read your content and didn't think it was worth keeping.
When you run run refresh recommender, Claude:
Pulls your Search Console coverage data
Flags pages that are decaying in rank or dropped from the index
Tells you whether to rewrite, merge, or kill each one
Optionally rewrites the page using System 2 so the new draft inherits your business context and the capsule method
This is the half of the job most people skip. Generating new content is only 50% of SEO. The other 50% is keeping the content you already have alive.
How do you put all four systems on a schedule?
You tell Claude Code to turn the workflow into a routine, and it sets it as a local automation.
The simplest version is one sentence:
“Set this workflow to run every Monday at 9:00 AM.”
Claude Code registers it as a routine. If you're using the desktop app, the catch is that your computer has to be on at run time. If you're using the CLI, you can run it as a cron job in headless mode, which is what I do across multiple sites.
If you want this fully cloud-based, you'd need to move the MCP connections (Search Console, DataForSEO) to the cloud too, which is more setup than most people want. Local cron is the 80/20.
And yes, before you ask, this all works in Codex with GPT-5.5 too. Same architecture, different runtime.
Frequently Asked Questions
Do I have to use Astro for this to work?
No. Astro just lets Claude publish posts directly without touching a CMS. WordPress, Webflow, Framer, all work, you just plug into their APIs or paste the MD file manually.
How much does DataForSEO cost to run all four systems?
The on-site audit was 48 cents per run on a small site. Keyword research is a few cents per cluster. Even if you ran the full stack weekly, you'd burn through the free $5 credit in a couple of months.
A blog structure where every H2 is a question and the first sentence answers it directly, so AI search engines can lift the answer cleanly. Full breakdown here.
Will the content writer trigger an AI penalty?
No. Google has publicly said AI content is fine when it's helpful. The reason this workflow doesn't trip penalties is the business context injection, the source citations, and the capsule structure. That's what “helpful” looks like.
Want me to set this up for you?
If you'd rather skip the wiring step and learn this inside a community of people running it in production, AI Ranking is where the full workflow lives. Live SEO audits every Thursday, weekly tutorials on systems like these, and a private repo of the agents, skills, and prompts I use across every site.
Connect Codex to Google Search Console and DataForSEO, and GPT-5.5 can run five SEO tasks for you on a schedule: a site health report, a click-through-rate rewrite agent, a content-idea agent, an internal-link agent, and a self-improving content engine. Two free API connections, five prompts, and your SEO does itself.
30 sec read
Skip to full article below
Why automate SEO with Codex in the first place?
Because the boring SEO tasks (the ones you actually skip) are the ones that move rankings. Pulling Search Console data, rewriting weak titles, finding pages with high impressions and zero clicks, hunting for content gaps. None of that is hard, it's just tedious. Codex with GPT-5.5 does it on a schedule and emails you the results.
I know I seem like I keep swapping from Claude to ChatGPT. Trust me, I don't want to give you shiny object syndrome, and I'm not telling you to jump ship to another AI for your SEO.
But if you're already using something in the OpenAI world, then you might want to know about this. Codex has a few things going for it right now (local automations, scheduling, GPT-5.5's tool use) that genuinely fit this workflow.
I'm here to help, I swear. Use whatever tool actually moves the needle for you.
What do you need before you start?
Two free things and one paid tool you probably already have.
A Google Cloud project with the Search Console API enabled (free, takes 90 seconds)
Codex running locally with GPT-5.5 selected (Codex high works too, but 5.5 is the sweet spot)
If you already have Ahrefs, you can skip DataForSEO and connect Ahrefs directly to Codex. Same outcome.
How do you connect Codex to Google Search Console?
You give Codex one OAuth credentials JSON file and it handles the rest.
Open Google Cloud Console, create a project, then go to APIs & Services.
Search for Google Search Console API and enable it.
Go to Credentials, create an OAuth client ID, choose Desktop app, name it "Codex", and download the JSON file.
In Codex, start a project in your site's folder and paste this prompt:
"Connect to the Google Search Console API using this OAuth credentials file. Authenticate me in the browser, save the token locally, then use the Search Console API to read my site's performance data."
A browser window pops up. Log in with the Google account that owns the Search Console property (this part trips people up: it has to be the right account). Click allow, and Codex saves the token. From now on, every chat in this project has live Search Console access.
How do you connect DataForSEO to Codex?
Grab two credentials and install the MCP server.
In your DataForSEO dashboard, go to API Access and copy your login email and password. If you've never logged in, the password is shown once at the top, save it. If you've logged in before, request a password reset and check your email.
Then ask Codex to install the DataForSEO MCP server with those credentials. Test it by asking Codex to pull ranked keywords for your domain. If you see a clean list back, you're good.
That's the whole setup. Two API connections, ten minutes, and Codex now has access to your Search Console performance and a full SEO dataset (keyword volumes, SERP data, Lighthouse scores, competitor intelligence).
A weekly on-page audit. Codex pulls Lighthouse data, checks for broken links, scores on-page SEO, and flags pages with high impressions but zero clicks. The first run on my Astro site came back with a 97 on-page score, no broken links, and a list of "pages to investigate" including a blog with average position 1.6 (basically ranking #1) but zero clicks. That's a title-tag problem, and Codex tells you exactly which one.
2. Click-through rewrite agent
Most people obsess over Google AI Overviews and forget that a strong title tag and meta description still drives the bulk of clicks. This agent finds pages with high impressions and weak click-through rates, runs the keyword through DataForSEO, analyzes the titles and meta descriptions of whoever ranks #1, and writes you better versions. Same impressions, more clicks, more traffic. It's the laziest win in SEO.
3. Content idea agent
Instead of generic keyword research, this agent reads your Search Console data, understands what your site is already winning at, then suggests new keywords that fit your business and have realistic difficulty. It outputs target keywords, suggested URLs, and the reason each one fits. This is the one that saves the most time, because thinking up topics is the bottleneck.
4. Internal link agent
The fourth prompt scans your site for internal linking opportunities. Codex finds pages on related topics that aren't linked to each other, then proposes the exact anchor text and source paragraph. It's the kind of work no one does manually, but it compounds rankings.
5. Self-improving content agent
The fifth one publishes blog posts. With your custom knowledge files loaded into the project, Codex can run on a schedule (Monday, Wednesday, whatever you set), pull a target keyword from agent #3, write the post, and save it to your repo. If you're on Astro, this is genuinely close to a self-writing website. WordPress users can do the same with a small wrapper.
How do you turn these agents into automations?
Once an agent's output looks right, ask Codex:
"Set this as an automation to run every Tuesday at 9:00 AM."
Codex registers it as a local automation. The catch: because it runs locally, your computer has to be on. I run mine Monday to Friday, 9 to 5, which matches when my computer is on anyway. If you want this fully cloud-based, it's possible but takes more setup.
You can also tell Codex to email you the output every run, which is what makes this actually useful. SEO reports you don't read are worse than no reports.
Frequently Asked Questions
Do I need to pay for Codex?
You need a ChatGPT subscription that includes Codex (Plus or higher). DataForSEO is pay-as-you-go and the free $5 credit covers a lot of testing.
Why GPT-5.5 specifically?
Codex high works, but 5.5 is consistently better at multi-step agent work and following long prompts without losing the thread. Use 5.5 unless you have a reason not to.
Can I use Claude or Gemini instead?
Yes, but Codex's local automations and tool-use are smoother right now. I find OpenAI's usage limits more workable than Claude's for this kind of repetitive agent work.
Will this work on WordPress?
Yes. The reporting and idea agents work on any site. The self-publishing agent needs a way to push to your CMS, which is straightforward with the WordPress REST API.
Is this the same as setting up an MCP server in Claude?
Conceptually yes. DataForSEO is an MCP server, and you're plugging it into Codex the same way you'd plug it into Claude. The difference is Codex has scheduled local automations baked in.
Want help setting this up?
If you're new to AI search SEO, the agents above will move faster with the right foundation underneath them. Inside AI Ranking, I teach the full workflow: from site structure to capsule content to getting cited by AI search engines. Live SEO audits every Thursday, weekly tutorials, and a community of people running this stuff in production.
A new DataForSEO study of 100,000 ChatGPT prompts found that 47% of them trigger fan-out queries: hidden searches ChatGPT runs against the web before it answers. If your content does not match those hidden queries, you do not get cited. Here is how to find them and write content that does.
30 sec readSkip to full article below
Why is there a "search you will never see" deciding if ChatGPT recommends you?
Because ChatGPT does not just take your question and go fetch an answer. It rewrites your question into multiple secondary questions, runs those against the web, then synthesizes the response. Those rewritten questions are called fan-out queries, and a DataForSEO study of 100,000 ChatGPT prompts found 47% of prompts trigger them.
Translation: the most important search query about your business is one no human ever typed. If your page does not answer that hidden query, ChatGPT skips you, even if you rank for the obvious keyword.
What are fan-out queries, exactly?
Fan-out queries are the secondary questions ChatGPT generates from your original prompt and runs against the web before answering you. They are the AI's way of saying "your query is not specific enough, let me search smarter."
Say a user types "best Italian restaurants in Chicago". ChatGPT might fan that out into:
best Italian restaurants in Chicago reviews
top Italian restaurant Chicago price
popular Italian restaurants Chicago menu
Each of those is a real search query running silently behind the scenes. The pages that match those fan-outs are the pages ChatGPT cites and recommends. This is the mechanic behind every modern AI search engine, and it is the core of generative engine optimization.
Why does matching fan-out queries equal citations?
Because citations are awarded to pages that answer the AI's hidden questions, not the user's original one. ChatGPT, Perplexity, and Google's AI Overviews all rely on fan-out style retrieval. Match the fan-out, get the citation. Miss it, get ignored.
This is why a page can rank #1 in classic Google and still never appear in ChatGPT search. The traditional "head term" mindset does not survive in AI search. You have to think in clusters of sub-questions, which is exactly what the capsule content method was built for.
Community win:William Moon, a financial advisor in Arizona, took his CTR from 0.3% to 2.3% by rewriting his pages to answer fan-out style sub-questions, then closed a $165,000 deal off one AI-driven lead.
How do you find the fan-out queries running for your niche?
You need a fan-out dataset. The fastest way is a free tool called DataWise, which pulls fan-out queries from DataForSEO's index and gives you 5 free uses on signup.
The flow inside DataWise:
Sign up free, land on the dashboard (your Google Search Console connection is optional).
Go to Keyword Research and pick the Fan-Out Queries tab.
Drop in your seed keyword, for example "local SEO", and hit Explore.
DataWise returns the fan-out queries triggered for that seed, plus AI search volume and trends.
Filter to only fan-out queries (you can skip "people also ask" and FAQs, which are different beasts).
Export the ones that make sense for you to a CSV. That is your content plan.
Heads up: as of recording, fan-out data is US-English only. The dataset is brand new, so expect international expansion soon.
If you are technical, you can also wire DataForSEO directly into Claude Code and run fan-out research from your terminal. I walk through that setup in this video.
Which fan-out queries should you actually answer?
Pick the ones that match your service, your audience, and your buyer intent. Not every fan-out is worth answering, and AI search volume is not the only signal. If a fan-out query makes sense for your business and the question is clearly being asked by ChatGPT, that is enough reason to write the page.
Quick triage rules:
Definitions: "what is X and how does it work" usually deserves a pillar post.
Verticals: "X for lawyers / dentists / agencies" become dedicated niche playbooks.
Comparisons: "X vs Y" or "is X worth it" map to comparison posts that AI loves to cite.
Process: "how to rank in Map Pack" become tactical how-to articles.
You do not need to answer all 99 fan-outs DataWise returns. Pick the 8 to 12 that fit your offer and start there.
How do you write content that actually gets cited?
You answer the fan-out query directly, in the first sentence of the section, then expand. AI engines are looking for self-contained, extractable answers, not 1,500 words of warm-up before you say anything useful.
The repeatable structure that works in 2026:
TL;DR or summary box at the top, 40 to 60 words, covering the whole article.
Every H2 phrased as a question that maps to a fan-out query.
Direct answer in the first 1 to 2 sentences of every section, then context.
Schema markup (Article, FAQ, HowTo) so engines can parse it cleanly.
Linked sources for every stat claim, because AI engines weight cited content higher.
A fan-out query is a secondary question ChatGPT generates from your original prompt and silently runs against the web before answering. It is how the model gathers enough source material to give a confident response. The DataForSEO 100,000-prompt study found 47% of ChatGPT prompts trigger fan-outs.
How is a fan-out query different from "people also ask"?
People also ask is a Google SERP feature populated from real human searches. Fan-out queries are AI-generated rewrites that no human ever typed, used internally by ChatGPT and other LLMs to retrieve sources. Some overlap exists, but fan-outs are uniquely a generative engine optimization (GEO) signal.
Do fan-out queries work outside the United States?
Not yet. As of April 2026, the DataForSEO fan-out dataset only covers US-English queries. International coverage is expected to roll out as the dataset matures.
Can I find fan-out queries for free?
Yes. DataWise gives you 5 free fan-out lookups when you sign up, no credit card. After that you can either upgrade or join AI Ranking School for unlimited access.
Will ChatGPT cite my page if I just stuff fan-out keywords in?
No, and this is where most people screw it up. You have to genuinely answer each fan-out query in a self-contained section with a direct opening answer, supporting context, and ideally a linked source. Keyword stuffing reads as spam to LLMs the same way it does to Google.
Want unlimited fan-out research and the playbook?
DataWise is free to try with 5 lookups. If you want unlimited fan-out queries, the full content writing system, weekly tutorials, and a community of operators implementing this every week, join AI Ranking School. That is where members like William, Tim Armstrong, and Steven (800+ pages, 105 appointments/month) are running this play.
