Physically based rendering describes surfaces in terms that correspond to how light actually behaves, so a material looks right under any lighting rather than being tuned for one scene.
Base colour (albedo)
The colour of the surface with no lighting information in it. No shadows, no highlights, no ambient occlusion — those come from the renderer.
The most common mistake is baking lighting into base colour, usually from a photograph. It looks fine under the lighting it was shot in and wrong under everything else.
Roughness
How scattered reflections are. Black is a mirror; white is completely diffuse. This is the channel that does the most work in making a material recognisable — the difference between polished and worn metal is almost entirely roughness.
Variation matters. A uniform roughness value reads as synthetic; real surfaces vary with wear, fingerprints, dust and moisture.
Metallic
Whether the surface is metal. It should be close to binary — near 1 for metal, near 0 for everything else. Mid-values are physically meaningless for a single surface and are only used at boundaries or where a mask blends.
Metals have coloured reflections and no diffuse component; non-metals have white reflections and a diffuse colour. That is why the flag matters so much: it switches how the whole surface is lit.
Normal
Perturbs the surface direction per pixel to fake detail. Adds the appearance of bumps, grooves and scratches without geometry — and cannot change the silhouette, which is why a normal-mapped surface gives itself away at a grazing angle.
Ambient occlusion
Where ambient light is blocked — crevices, contact points, tight corners. Adds contact grounding. Should be subtle; heavy AO reads as dirt.
Emissive
Light the surface produces. Screens, lamps, hot metal, glowing runes. Black everywhere else.
The metallic/roughness versus specular/glossiness split
Two conventions exist. Metallic/roughness is the more common in real time and what glTF standardises on. Specular/glossiness carries slightly different information. Converting between them is possible and lossy; picking one per project avoids the question.